Is newer better? Penn World Table Revisions and their
impact on growth estimates
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation
Johnson, Simon, William Larson, Chris Papageorgiou, and
Arvind Subramanian. “Is Newer Better? Penn World Table
Revisions and Their Impact on Growth Estimates.” Journal of
Monetary Economics 60, no. 2 (March 2013): 255–274.
As Published
http://dx.doi.org/10.1016/j.jmoneco.2012.10.022
Publisher
Elsevier B.V.
Version
Author's final manuscript
Accessed
Thu May 26 00:28:17 EDT 2016
Citable Link
http://hdl.handle.net/1721.1/87736
Terms of Use
Creative Commons Attribution-Noncommercial-Share Alike
Detailed Terms
http://creativecommons.org/licenses/by-nc-sa/4.0/
Is Newer Better?
Penn World Table Revisions and Their Impact on
Growth Estimates∗
Simon Johnson†
William Larson‡
Chris Papageorgiou§
Arvind Subramanian¶
June 2010
Abstract
This paper sheds light on two problems in the Penn World Table (PWT) GDP estimates. First,
we show that these estimates vary substantially across different versions of the PWT despite
being derived from very similar underlying data and using almost identical methodologies; that
this variability is systematic; and that it is intrinsic to the methodology deployed by the PWT to
estimate growth rates. Moreover, this variability matters for the cross-country growth literature.
While growth studies that use low frequency data remain robust to data revisions, studies that
use annual data are less robust. Second, the PWT methodology leads to GDP estimates that are
not valued at purchasing power parity (PPP) prices. This is surprising because the raison d’être
of the PWT is to adjust national estimates of GDP by valuing output at common international
(purchasing power parity [PPP]) prices so that the resulting PPP-adjusted estimates of GDP
are comparable across countries. We propose an approach to address these two problems of
variability and valuation.
JEL Classification: O11, O40, O47.
Keywords: Penn World Table, GDP estimates, purchasing power parity, revisions, methodology, growth.
∗
Special thanks go to Angus Deaton, Alan Heston, David Romer, and David Weil, for comments, encouragement
and suggestions. We also benefited from discussions with Daron Acemoglu, Craig Burnside, Francesco Caselli, Michael
Clemens, Erwin Diewert, Robert Feenstra, Benjamin Jones, Chad Jones, Aart Kraay, Benjamin Olken, Valerie Ramey,
Mick Silver, Antonio Spilimbergo, Nancy Stokey, Alwyn Young and seminar participants at Brown University, the
Bureau of Economic Analysis, IMF, Oxford University, the University of Pennsylvania (2008 Penn World Table
workshop), Princeton University, Tilburg University, the World Bank, the Econometric Society Summer Meeting
(Boston 2009), the NBER Summer Institute (Economic Growth workshop, and Wealth and Inequality workshop
2009), the Society for Economic Dynamics Meeting (Istanbul 2009), and the conference on Measuring and Analyzing
Economic Development, (University of Chicago, February 2010). The views expressed here are those of the authors.
†
MIT Sloan School of Management, Peterson Institute for International Economics, and NBER; email: sjohnson@mit.edu.
‡
George Washington University; email: larsonwd@gmail.com.
§
IMF; email: cpapageorgiou@imf.org.
¶
Peterson Institute for International Economics, Center for Global Development, and Johns Hopkins University;
email: asubramanian@piie.com.
1
1
Introduction
How fast did Equatorial Guinea grow over the two and a half decades beginning in 1975? The
natural place to turn to answer such a question is data from the Penn World Table (PWT), which
is the most widely used source for cross-country comparisons for the level and growth rate of GDP.1
According to its latest available version (PWT 6.2) Equatorial Guinea is the second-fastest growing
country among 40 African countries. However, according to its previous version (PWT 6.1), which
was released four years before, Equatorial Guinea was the slowest growing country. Indeed, as table
1 shows, if one were to compile the list of the 10 fastest and slowest growing countries in Africa
between 1975 and 1999, PWT 6.1 and PWT 6.2 would produce almost disjoint lists. Both would
agree that Botswana has done best, and they would also agree that Egypt, Cape Verde, Lesotho,
Mauritius, Morocco, and Tunisia belong in the list of fast growing countries. But there are six
countries that are not identified by both versions as fast growing. The disagreement is even more
severe for the “worst performers.” There are a total of 10 countries that appear to have the slowest
growth according to one version but not according to the other. Such data variability is especially
surprising because PWT 6.1 and 6.2 use very similar underlying data and methodologies. The
concerns about GDP data quality and the variability it engenders have recently led researchers to
explore alternatives to GDP data, and Henderson, Storeygard, and Weil (2009) and Young (2009)
are noteworthy examples.
Before we examine this problem of data variability, we need to describe briefly what the PWT
was designed to do. The pioneering work of Irving Kravis, Alan Heston, and Robert Summers,
which led to the Penn World Table data, was aimed at converting national measures of GDP
and income into internationally comparable estimates. Cross-country comparisons could not be
based on national GDP data because these were valued at domestic prices. Since some goods and
especially services were known to be cheaper in poor countries compared to rich countries, adjustments needed to be made to the valuation of these goods and services so that they could be made
internationally comparable. These adjustments were made by calculating common international
prices–the so-called purchasing power parity (PPP) prices–for all goods and services. With these
PPP adjustments, GDP could then be compared across countries.
1
Roughly two-thirds of all cross-country empirical work is based on PWT. Second place is held by the World Bank’s
World Development Indicators (WDI), which were originally based on the PWT but have subsequently diverged. The
IMF’s World Economic Outlook (WEO) dataset places a distant third.
2
A large literature has assessed the basic methodology employed by the PWT for determining
these PPPs. In addition to the series of papers by Kravis, Heston, and Summers (1978) and
Summers and Heston (1980, 1991, and 1996), notable contributions include Ciccone and Jarocinski
(2010), Deaton (2006), Dowrick (2005), Dowrick and Quiggin (1997), Feenstra, Heston, Timmer and
Deng (2009), Heston (1994), Neary (2004), Nuxoll (1994), Rao and Selvanathan (1992), Samuelson
(1994), Srinivasan (1994), and van Veelen (2002).
However, much of the discussion and criticism of the PWT methodology, including most recently
by Deaton and Heston (2010), has focused on the PPP and GDP estimates for the benchmark year
for which disaggregated data are collected from the countries. Only recently has attention been
given to the intertemporal dimension of the PWT and its methodology, namely its estimates for
the growth rates of PPP-adjusted GDP, which then has implications for PPP-adjusted GDP levels
for nonbenchmark years. Feenstra, Heston, Timmer and Deng (2009) focus on the problems for
PWT created by revisions to net exports. Feenstra, Ma, and Rao (2010) derive theory-consistent
intertemporal data for consumption (but not for GDP) while Rao, Rambaldi and Doran (2009) use
an econometric estimation procedure for deriving intertemporal estimates for the PPPs and GDP.
Most notably Katayama and Ponomareva (2010) were the first to explore the variability of
growth rates resulting from revisions in PWT. More precisely, this paper documents how growth
rates change across four versions of the PWT — 5.0, 5.5, 5.6 and 6.1 — and consequently analyzes the
effects of these revisions on the main empirical finding of a negative relationship between volatility
and growth in Ramey and Ramey (1995). Katayama and Ponomareva report that across the four
different versions of PWT, 13 percent of their sample of 110 countries show a sign change in their
annual average growth rates for the period 1962-1985. They also show that the Ramey and Ramey
finding supported in some versions of the PWT but not in others.
The focus of this paper, as in Katayama and Ponomareva (2010), is the time dimension of
the PWT, specifically the data revisions across versions of the PWT. Unlike Katayama and Ponomareva, however, we concentrate on the sources of variation in growth rates between versions 6.1
and 6.2 for which the underlying price data used for the construction of PPP-adjusted GDP growth
rates are the same. Therefore, the main focus of our paper is the PWT methodology for constructing growth rates of GDP, and the resulting estimates for the levels of GDP for nonbenchmark years.
It draws attention to two problems–variability and valuation–in the PWT’s estimates of GDP
(growth and level) and PPPs.
3
Our summary findings are: First, data revisions are substantial. This is particularly surprising
because, as we describe below, versions 6.1 and 6.2 differ in relatively small ways in terms of data
and methodology. Second, revisions are systematic and inherent to the PWT’s methodology for
computing growth rates and estimates for the nonbenchmark years. Revisions are more pronounced
at high frequency (annual data are much more variable than longer averages); for small countries
(i.e., countries with small total GDP); and for historical data, so that the further the data are from
the benchmark year in the PWT, the more variable they are.
Third, the variability of growth data has implications for the cross-country growth literature.
Results are indeed robust in the long-run (30-year averages) across versions of PWT which is a
positive and hopeful result, given the magnitude of both PWT and the ICP and related literature.
However, results based on annual data prove to be less robust across versions of the PWT than
are results based on 10-year averages and/or levels of GDP. Results are also sensitive to sample,
especially the inclusion of small countries.
Fourth, the PWT methodology raises a more basic question about valuation. The rationale for
the PWT is to come up with GDP level and growth data that are at common international (the
so-called PPP) prices so that the data are comparable across countries. The methodology, however,
leads to the construction of GDP growth estimates that are based not on common international
prices but on a mixture of international and domestic prices; as a result GDP level estimates for
years other than the benchmark year are also not at international or PPP prices. In this case, it is
not obvious that the data are comparable across countries.
Finally, we propose an alternative way of using the PWT data that might address the problems
of both variability and valuation. Essentially, we urge the construction of a new chained series,
with all data valued at common international prices, and based on greater use of the disaggregated
data collected for the different benchmark years.
This paper is organized as follows: Section 2 briefly reviews the background and history of
the PWT and illustrates key aspects of the methodology used in the PWT calculations of GDP
growth rates. Section 3 shows the variability of growth data and the underlying patterns to this
variability, focusing on versions 6.2 and 6.1 of the table. In section 4, we summarize the results of our
robustness/replication studies for leading growth studies. In section 5, we explain why not all PWT
data fully reflect PPP prices. Section 6 describes our proposal that addresses the problems with
PWT methodology relating to data variability and to data not being at PPP prices. Section 7 draws
4
some conclusions, including offering some practical suggestions for researchers. Technical appendix
1 contains methodological details on the construction of the PWT data. Technical appendix 2
illustrates the valuation problem, namely why the current PWT methodology leads to GDP level
and growth estimates that are not at international prices.
2
Penn World Table: Background and Methodology
A. Background and history
The PWT has reported on seven rounds of data, starting in 1970 (table 2). The latest published
version is 6.2, which was released in October 2006.2 The prior version was 6.1 and that was preceded
by 5.6. These are the most commonly used versions of the PWT today, and it is on these that we
focus our attention. However, it is our understanding that the points made below also likely apply
to other versions.
The core purpose of the PWT is to produce measures of real GDP from the corresponding
relative price levels across countries and over time based on prices for the same or similar goods
in different countries. The price collection operation, which is at the heart of PWT, is known as
the International Comparison Programme/Project (ICP).3 Each “generation” of the table is based
on a different round of the ICP; so, for example, versions 5.6 and 6.1 use different prices. The
benchmark year of the PWT corresponds to the year for which the ICP exercise is carried out. The
number of countries participating in the ICP has increased steadily, starting from 10 countries in
1970 to 146 countries in 2005. From table 2, it can be seen that the number of countries for which
GDP estimates are provided by the PWT exceeds, often considerably, the number of countries
participating in the ICP. Within a generation of the PWT, the ICP remains the same, e.g., this is
the case for 6.2 and 6.1, but there are other revisions–the nature of which varies.
The ICP is a massive undertaking, requiring a vast amount of resources–mostly in terms of
people’s time, but the computer resources, at least until recently, were also significant. The only
organizations that can sustain such an investment are government-funded, and even international
organizations, such as the United Nations, have a hard time coming up with all the money required.4
2
Version 7.0 will be available soon.
The International Comparison Project began in 1968, although its antecedents date back to the 1950s. Irving
Kravis was the first director; http://unstats.un.org/unsd/methods/icp/ipco_htm.htm. In 1989 the “P” became
Programme, rather than Project. International price comparisons have been completed for 1970, 1975, 1980, 1985,
1990, 1996, and 2005.
4
The Ford Foundation did provide critical early financial support through grants to the University of Pennsylvania.
3
5
A substantial amount of the needed investment has come directly from governments.
As a result, control of the ICP has shifted over time. “After phase III [1975], the role of the
University of Pennsylvania, which had until then been the main engine of the ICP, was gradually
transformed into that of adviser on methodological issues. Another notable change in ICP responsibility was the increasing role of the Statistical Office of the European Communities (Eurostat).
Eurostat, in fact, became not only the organizer of the European Community comparison, but also,
with its experienced staff, it has provided substantial technical assistance to a number of regional
comparisons and to the work on establishing links among the various regions.”5
As control has shifted, so have preferences regarding methodology both for particular regions
and for how these are aggregated to global estimates. The most important change may have been
regionalization: “In phase IV [1980] and onward, countries participated through regions or country
groups; first regional (e.g., African, OECD, etc.) comparisons were carried out and then the world
comparison was built up by linking across these groups.”6 Deaton and Heston (2010) focus on
problematic methodological points, all of which arise because particular organizations control part
of the methods. Their concern suggests that the methodology changes significantly such that it is
harder to say if the PWT is improving or not in its accuracy. But there are other reasons to worry
about this exact same issue.
B. Methodology
The innovation and great contribution of the PWT was to convert national measures of GDP and
income into internationally comparable PPP estimates. This is done–in principle–by collecting
prices for the same or similar goods in different countries and deriving price indices that can be
used to compare what people can actually buy. The PWT obtains local currency data from the
national income accounts of countries. Then, based on international price comparisons, it converts
these local currency data into PPP-based figures, which should be comparable across countries.
But how are estimates derived in practice by the PWT? The PWT has two distinct approaches
and steps for calculating the estimates of domestic absorption and its constituents and GDP. The
first step (with a distinct methodology) relates to estimates for the benchmark year. And the
second step (involving a different methodology) relates to estimates for growth and to the level
estimates for all nonbenchmark years.
5
6
Quotation is from http://unstats.un.org/unsd/methods/icp/ipco_htm.htm.
Ibid.
6
We describe both steps of the PWT methodology in detail in technical appendix 1. In this
paper we focus on the second step, which involves estimation of GDP growth and hence the level
of GDP in nonbenchmark years.
Once the PWT calculates the level of PPP-adjusted GDP (and the associated domestic absorption DA, which is the sum of consumption, investment and government expenditures) for the
benchmark year (say 1996), it calculates the levels for nonbenchmark years according to the following equations. GDP for a nonbenchmark year, say 1995, is calculated as:
95 = 95 +   95
(1)
where Y is PPP-adjusted GDP, DA is domestic absorption and NFB is net foreign balance (which
is the difference between exports and imports of goods and nonfactor services). By definition,
´
³
d 9596
95 = 96  1 + 
(2)
where the hat sign over a variable denotes growth rate.
Now 96 is domestic absorption for the benchmark year which is, by definition, in PPP terms.
d 9596 is computed:
Consider next how 
d 9596 =  ̂9596 +  ˆ9596 +  ̂9596 

(3)
where  ,  ,  are the shares of consumption, investment, and government spending, respectively, in domestic absorption.
The growth rate of domestic absorption is calculated as the weighted average growth rate of its
three components, C, I, and G. The weights assigned to each of these components are the shares
of each component in domestic absorption measured in 1995 and measured at international prices.
These shares are obtained from step 1 described in technical appendix 1. It should be stressed,
however, that the growth rates of real C, I, and G, are from the national income accounts (constant
price series), so that they are by definition computed at domestic prices and not at international
prices, a key issue that we will explore subsequently.
In sum, the PWT calculates growth rates of DA and GDP using equation 3 above and uses
these growth rates to derive estimates for the level of GDP in nonbenchmark years. One important
corollary of this procedure is worth highlighting. Take versions 6.1 and 5.6 of the PWT, which had
1996 and 1985, respectively, as the benchmark years. In PWT 6.1 the estimate for the level of per
7
capita GDP for 1985 is derived from the level estimate for 1996 (the benchmark year for PWT 6.1)
to which the growth rate, as calculated in equation 3, is applied. The information on PPP-adjusted
GDP per capita in PWT 5.6, based on the disaggregated data compiled in the ICP for 1985, is
only partially used. To be precise, previous benchmark data are used in the calculation of the level
of GDP in the subsequent benchmark year: for example, PWT 6.1 incorporated some of the 1985
benchmark year inputs (in PWT 5.6) into the 1996 benchmark year calculations.7 This occurs
for every new version so that potentially valuable information from previous ICPs is not used. Of
course this is at the heart of the problem of reconciling benchmark studies with national accounts
data, explored in early work by Summers and Heston (e.g. Summers and Heston, 1984), in what
was described in the 1980s as the ‘consistentization’ problem and more recent studies by Eurostat
and the OECD on reconciling their annual benchmarks and national accounts growth rates.8
3
Data Variability: Patterns and Associations
We focus on the variability of estimates between PWT version 6.1 and 6.2. These versions differ
in some relatively small ways.9 Both versions are within the same generation of the Table, and
therefore have almost identical methodologies and, more importantly, between these versions there
is no new ICP and hence no new international price data.10 Despite these small changes, estimates
of GDP levels and growth vary substantially between PWT 6.1 and 6.2. Of course, one substantive
difference is that PWT 6.2 uses more “updated” national income accounts data than PWT 6.1,
but in practice this updating does not lead to large revisions.11
Figure 1 illustrates that the 29-year average annual growth rates differ between PWT 6.2 and
7
For a full description see page 9 of the Data Appendix http://pwt.econ.upenn.edu/Documentation/append61.pdf.
Although consistentization did reconcile national growth rates with implied growth rates of multiple benchmarks
by using information of past benchmarks and current benchmarks, it proved difficult to implement. For details see
the unpublished appendix to PWT 5.
9
There are new PPPs for the OECD countries, 20 additional countries (mainly transition economies), coverage
was extended from 2000 to 2004, and the reference year for calculation of the purchasing power parities has been
moved from 1996 to 2000.
10
In addition, an important point emerges from a careful examination of the methodological history of the tables
(e.g., Deaton and Heston 2008). We usually think that a revised data series is better than the original series due to
various kinds of corrections. But in the case of the tables, the methodology has not necessarily improved over time.
There have been innovations, but some of these happened for bureaucratic or even political reasons as international
organizations became involved in the data collection, preparation, and presentation. As a result, it is not reasonable
to assume that one version of the tables (e.g., the latest) is necessarily better than other versions. This is another
reason why results that hold in one version of the tables but not other versions should probably be viewed with
greater skepticism.
11
The correlation between the 29-year annual average growth rate in the national income accounts data used in
the two versions is 0.99.
8
8
6.1.12 The average difference in the growth rate (for the period 1970—1999) generated by the two
versions is close to zero (0.1 percent). But the standard deviation of the differences in growth is
about 1.1 percent. This is quite substantial when compared with the fact that the average growth
rate (in 6.2) is 1.56 percent. Put differently, growth is more than 1 percent per annum different in
more than half the countries and more than 2 percent different nearly a quarter of all cases.
Since the PPPs are critical to the PWT methodology, we show in figure 2 the variation in the
estimates of the level of the PPPs (averaged over 29 years). There is a great deal of variation across
versions in the calculated change in prices, measured in PPP terms.
One basic aspect of data variability between PWT versions–namely that it increases when the
data are at higher frequency–becomes evident when we compute the growth rate over 1-, 10-, and
29-year periods. These are presented in the three panels of figure 3 (where the third panel is the
same as figure 1) with the same scale for the growth rates. Figure 3 dramatically illustrates that
growth computed on an annual horizon is considerably more variable across versions than growth
computed over 10- and 29-year horizons. For example, the standard deviation of the growth rates
across the two versions is 5.39 percent (relative to the average growth rate of 1.86 percent over
that horizon) compared with a standard deviation of 1.08 percent (relative to average growth of
1.51 percent) for the 29-year horizon. The striking differences in this figure intuitively explain the
results we find in the next section about the robustness of leading growth studies. Results based on
annual data prove to be less robust across versions of the PWT than are results based on 10-year
or 29-year averages. Evidently, the errors we are seeing get averaged out over longer horizons.
Figure 4 shows some prominent cases of revisions in the GDP numbers across the three latest
PWT versions. This affects many of the narratives about growth and the associated policy discussion for some low income countries stemming from the very basic question of who has done well
and who has done badly, say over the past 25 years.13 Table 1 illustrated the severity of the data
problem for Africa.14
12
This 29-year period is the longest for which the sample size can be maximized. Unless otherwise specified, all the
growth rates are for the RGDPCH series in the PWT. This is a chained series and is the one endorsed by the authors.
In technical appendix 1, we explain briefly the difference between this series and the RGDPL series. Appendix table
1 lists the countries in the sample. The distribution of the differences shown in Figure 1 is broadly similar to the
distribution of the differences found when comparing PWT 6.2 with PWT 5.6 (see table 6), WDI, or WEO data.
13
This basic and important question was asked, for example, by the World Bank’s Growth Commission, headed by
Michael Spence.
14
More generally, do the leading alternative versions of growth data of the tables agree on the fastest growing
countries around the world? For this we can compare “top 10” and “bottom 10” in terms of growth performance
according to versions 6.1 and 6.2 of the tables, alongside growth rates calculated from WEO and WDI. All four sources
9
A. Associations of variability in GDP estimates across versions
Next, we examine systematically–based on simple regressions–the associations of data variability in the PWT. Our aim is not to establish causality but to understand the possible influences
on data variability. In this spirit, we run regressions for three different and important variables
estimated by the PWT–the PPP prices, the level of PPP GDP, and the growth rate of PPP GDP.
We need to identify the possible influences. Why should the PWT estimates vary across revisions? Four possible factors may be at work. First, and most obviously, PWT estimates could
change because the underlying National Income Accounts (NIA) data–which are key inputs for
the PWT system–change.
Second, even apart from changes in the underlying data, revisions could be systematically related
to the quality of the data. The PWT has always been quite transparent about this data quality
issue, and it prominently assigned “quality grades” for each country (e.g., to version 5.0). These
grades are subjective assessments made by the authors of the PWT, based on a number of factors
described in the technical appendix to PWT 6.1, pages 13—18. Figure 5 shows the breakdown of
countries by their grades in version 6.1. Strikingly, only 32 countries received a grade of A or B.
Grades of C or D were received by 147 countries. Plotting the data suggests that data quality
might matter for revisions. The left-hand panel in figure 6 shows differences in 29-year annual
average growth rates (1970—1999) for countries with data quality grades of A or B. The right-hand
panel shows the same for countries with grades of C or D. All the major variation across versions of
the table occurs in the countries with lower grades. The same comparison also holds across higher
frequency growth data.
Two other factors–size of a country and time–are potentially important and suggested by the
PWT methodology for calculating the PPPs. The importance of size in affecting data variability was
formally highlighted by Rao and Selvanathan (1992). They show that the PPPs and international
prices can be seen as weighted averages, which makes it possible to interpret them as estimators of
parameters from appropriately specified regression models. It then also becomes possible to assess
the reliability of the estimates of these parameters. They show that the standard error of estimates
agree on seven of the fastest growing countries: South Korea, China, Botswana, Thailand, Hong Kong, Ireland, and
Malaysia. But there are also seven countries that appear on at least one list but not on all lists. And even between
the IMF and the World Bank data there is disagreement on the exact ranking (and on whether Indonesia, Portugal,
and Luxembourg make the cut). Again, there is more disagreement on the countries that have done worst–probably
because poorer performing countries tend to have less reliable data. Only five countries appear on all four lists:
Madagascar, Nicaragua, Togo, Venezuela, and Zambia.
10
of the PPPs are inversely related to a country’s size (more specifically, its total consumption
expenditure).15 A proposition that we will test in the data is: the smaller the country, the less
reliable is the estimate of its PPP.
The final factor is time.16 The GDP estimates for the nonbenchmark years in the PWT are
based on extrapolated PPPs. The methodology works as follows. Take 1995, the first nonbenchmark
year in PWT 6.1. The calculations of the international prices for 1995 are estimates based on the
actual price data for 1996. So, some error is added to the 1995 estimates. If we take this back one
more year to 1994, we know that 1994 international prices are in turn estimated from 1995 numbers.
This adds one more layer of error. Relative to the benchmark year 1996, there are two sources of
error for the 1994 estimates, and so on. This error structure going back in time is analogous to
the error structure going forward imposed by many forecasting exercises. It might therefore be
expected that estimates are more variable farther away from the benchmark year–a proposition
we test in the data.17
B. Regressions results
Regressions for revisions to the PPPs, level of GDP, and GDP growth are reported in tables
3—5 respectively. The first variable we need to measure is the change in the underlying national
income accounts data across versions. But there is a complication here that merits explanation.
In the PWT, the underlying national income accounts growth data change for two reasons. First,
national authorities can, and occasionally do, revise data. To capture the revisions done by the
national authorities, we take GDP growth data from the World Development Indicators (which
reflect data submitted by authorities) for the years corresponding to PWT 6.1 and 6.2, which are
2002 and 2005, respectively. We use the log of the absolute value of differences in the growth
of GDP (at constant domestic prices) between these two datasets as our measure of changes in
national income accounts data due to national authorities’ revisions. But there is a second source
of change to national income accounts data emanating from the PWT procedures. That is, the
PWT further revises the data submitted by the country authorities before using it for computing
15
See equation 5 in Rao and Selvanathan (1992).
We should note that it is not the case that revisions to the table change only more recent data. Appendix figure 2
shows that there are large revisions for the 1970s, the 1980s, and the 1990s looking at 10-year average annual growth
rates.
17
Rao and Selvanathan (1992) show that that the standard error of the Genebralized Least Squares (GLS) estimator
for the PPPs is not only related to the size of a country (noted above) but also to the variance of the error term.
Since, the PPPs are extrapolated sequentially, random error is added at each stage, so that the variance increases as
we move away from the benchmark year.
16
11
PPP-based estimates described in technical appendix 1.18 We use the log of the absolute value of
differences in the growth of GDP (at constant domestic prices) between these two PWT-generated
datasets as our measure of changes in national income accounts data due to the PWT revisions.19
We use two alternative measures for data quality. The first is the grading by PWT itself. We
convert the four letter grades for data quality into an index, from one to four, where four is the best
(A). The second measure is simply the number of ICP studies in which a country has participated.
For size, we use the log of a country’s PPP-adjusted total GDP obtained from PWT6.2. Since the
benchmark year for PWT 6.1 is 1996, our measure of time is the number of decades from 1996
(with each year being 1/10 of a decade, the datapoint for 1995 will therefore be 0.1).
So, the basic regression we run is:


 (   ) = 0 + 1  (
) + 2  (
) +  1    +
 2  +  3  +  4  +  5 ( ∗  ) +  
(4)
where  () refers to the absolute value of the difference in the estimate between PWT 6.2 and 6.1
of the relevant variable within brackets;  and  subscripts refer to country and time, respectively.
 is either the log of PPP prices (in Table 3), the log per capita PPP-adjusted GDP (in Table
4), or the growth rate of per capita PPP-adjusted GDP (in Table 5).    is the growth rate of
per capita GDP in constant domestic (i.e., not PPP-adjusted) prices as reported by the national
income authorities;     is the growth rate of per capita GDP in constant domestic (i.e., not
PPP-adjusted) prices as revised by the PWT;    is the measure of country data quality
as evaluated by the PWT.  refers to the number of ICP studies in which a country has
participated.  is PPP-adjusted total GDP and is obtained from PWT6.2.  is calculated
as the absolute difference between the year of the observation and 1996 (i.e., abs(t-1996)), and is
divided by 10 to express the variable in decades.  ∗  is the interaction between the distance
variable and the number of ICPs variable, and is a random error term.
In table 3, for example, the dependent variable is the absolute value of the difference in log
of international price between version 6.2 and version 6.1.20 Columns 1 and 2 show that the two
18
This layer of revision by the PWT appears to involve merging and splicing national income accounts data from
multiple sources to derive a long series and appears to be done for low-income countries.
19
In tables 3-5, we call the first source of change, “NIA-GDP growth” and the second source of change “PWT-NIA
GDP growth.”
20
All the regressions are done at annual frequency but we have also replicated them at decadal and 25-year
frequencies and obtained similar results with regards to coefficient signs and significance levels. Consistent with
figure 3, when using higher frequency data the coefficient magnitudes are considerably larger for all regressors.
12
sources of changes in the national income accounts data do affect revisions to the PPP. In column
3, it can be seen that a higher grade is negatively related to differences in the data, which confirms
that the PWT team warnings on data quality were appropriate. Column 4 shows that if there
have been more ICP benchmark studies for a country, this also reduces the extent of data revisions
between versions 6.2 and 6.1. Log GDP (column 5) has a similar effect, although this may be a
proxy for data quality or the existence of a benchmark study. These results are consistent with the
quality grading as advocated for years by the PWT researchers.
Perhaps more novel is the result in column 6. This shows that the further a data point is from
the benchmark year, measured in absolute value of years from 1996, the more likely its data is to
change between versions 6.1 and 6.2, i.e., the coefficient is positively signed. These results on the
variability of the PPPs and their relationship to country size (log total GDP) and distance are
graphically represented in figures 2 and 7, respectively. Figure 7 is striking in depicting a funnel
relationship between variability and distance from benchmark year, particularly for countries with
grades C and D. Data variability increases as the data point moves away from 1996 both forward
and backward in time. A key implication is that historical data in the PWT need to be viewed
with particular caution. When we combine all our explanatory variables (as in column 7), all of
them, except the “NIA-GDP growth” variable, continue to be significant. That is, the revisions
to the PPP computed by the PWT are explained by the changes to the national income accounts
data made by the PWT as well as other aspects of PWT methodology.
Table 4, which reports results for the level of GDP per capita, suggests that the determinants
of the revisions for GDP are similar to those for international prices.
Table 5, which explains revisions to the growth of GDP per capita, shows similar results with
one exception. It turns out the revisions to the PPP-adjusted growth rates computed by the PWT
are largely determined by the revisions to the underlying national income accounts data. These
revisions matter partly due to revisions done by the country authorities as illustrated in column
7 by the highly significant coefficient on “NIA GDP growth,” but also due to revisions by the
PWT procedures illustrated in column 8, where the variable “PWT-NIA GDP growth” trumps the
variable “NIA GDP growth.” Other aspects of the PWT methodology such as data quality and
country size remain significant determinants of PPP-based growth revisions.21
21
We have also reproduced tables 3-5 for the subsamples of countries with data quality A&B and C&D, respectively.
As expected, most of the variation in the data comes from the C&D subsample. Notably, the  coefficient estimate is
large, positive and highly significant in the C&D variable, while close to zero and insignificant in the A&B countries.
13
4
Replication Exercises for the Cross-Country Growth Literature
Thus far, we have shown that PWT data vary across revisions and do so systematically. But do
they matter? To assess this we turn now to examine some of the most prominent studies in the
growth literature that have used PWT data.
Most papers in the empirical growth literature use as a dependent variable either the level of
real per capita GDP in PPP terms or the growth rate in the same. Some papers use the level of
income or growth rates as a right-hand side control variable. Given the considerable variation in
the data on growth rates (and levels) and the fact that there is no one “best” set of PWT to use,
a natural question is: Which results in the literature are robust to changing versions of the table?
A. Criteria for considering papers
We have examined many of the leading papers in the growth literature based on PWT 5.6 or
6.1.22 In each case, we attempted to run exactly the same specifications and samples, but using
version 6.2 of the table instead. This approach cannot prove that a particular set of results is right
or wrong, but it may illustrate patterns in terms of what kind of results are more or less robust.
We had four criteria for inclusion. First, the list had to include papers that studied the level of
GDP, the growth rate of GDP, and the volatility of GDP. Second, since data revisions can be thought
of as measurement error, and since estimation depends upon whether measurement error is on the
left- or right-hand side, we included papers in which the various GDP measures (level, growth, and
volatility) featured as regressors and as regressands. Third, since our preliminary analysis of the
data showed that measurement error varies significantly between high and low frequency data, we
chose papers that adopted a pure cross-sectional approach (very low frequency); a panel approach,
with data averaged over four to five years (high frequency); and a panel approach with annual
data (very high frequency). Appendix table 2 lists the various papers in the appropriate analytical
category. After having decided the universe of potential papers based on these criteria, we narrowed
the list to papers that we considered influential. Finally, for inclusion here, we had to be able to
obtain the original data and to be able to replicate the paper’s main results with those data.
Appendix figure 3 depicts the relationship between growth variability and distance from benchmark year. The
complete set of regressions results is available upon request.
22
We focused our attention on “high impact” papers, measured either in terms of citations (using Google Scholar)
or based on discussion with active researchers or papers that we think will prove influential. We sought to examine a
range of papers, in terms of the frequency of data and methodology. However, our sample is not comprehensive, and
the results are intended as illustrations only.
14
In the replication exercise, we first replicate the authors’ core result based on the original data.
Next we show the same core results using the original data but changing the sample to match what
is available in PWT 6.2. We then report the core result using PWT 6.2 data (holding the sample
constant). Finally, we reproduce results for a more truncated sample that excludes countries with
a data quality rating of D.
B. Papers with table invariant results
In all, we tested the robustness of 13 papers in the growth literature. Note that we did not
check all specifications in all papers. Rather we concentrated on what appeared to us–or to others
citing the work–as the “main” results. The lower part of Appendix table 2 lists nine papers for
which we found basically no or small changes in results. In addition, there were more substantial
changes for four papers: Ramey and Ramey (1995), Jones and Olken (2005), Hausmann, Pritchett,
and Rodrik (2005), and Aghion, Howitt, and Mayer-Foulkes (2005). We go through these in more
detail below.
Before that, we should emphasize that many prominent papers were unaffected by the robustness
checks. Whatever else is right or wrong with the growth literature, the bulk of it is not afflicted
by the problem of sensitivity to changes in PWT GDP data. This list includes work such as
Acemoglu et al. (2003), Barro (1999), Burnside and Dollar (2000), DeLong and Summers (1991),
Demirguc-Kunt, Laeven, and Levine (2004), Easterly et al. (1993), and Sachs and Warner (1995).
The minor changes that we found in results from these papers are mentioned briefly in Appendix
table 2. Miguel, Satyanath, and Sergenti (2004) proved robust in a particularly interesting way,
which we expand on below. And we also discuss Mankiw, Romer, and Weil (1992) in more detail
below because their results come close to being affected despite involving level GDP regressions
that proved very robust for other papers.
C. Papers with changing results
Ramey and Ramey (1995 )
This paper, published in the American Economic Review in 1995, tests the link between growth
and volatility. The specification involves running an annual panel with growth rate of per capita
GDP on the left-hand side, a set of country- and time-varying controls on the right-hand side, and
a time fixed effect.23 There is no country fixed effect; instead there is a country-specific and time23
The equation estimated is: ∆ =   +  +  , where  is the log level of per capita GDP,  is the standard
15
invariant measure of volatility of growth, proxied by the standard deviation of the country-specific
residuals over the period covered by the growth data. Residuals and the volatility measure are
simultaneously estimated using maximum likelihood.
The paper used growth data from the PWT 5.6. In the original paper, there are 2,208 observations from 92 countries, while in the balanced sample (i.e., where there are data from both PWT
5.6 and 6.2) there are 1,776 observations from 74 countries.
In the paper, the coefficient on volatility is negative and significant at the 1 percent level. The
magnitude of the coefficient is -0.177. We replicate this result in the first column of table 7. When
we re-estimated their core specification on the balanced sample (column 2), the coefficient on the
volatility term is smaller and barely significant. When we switch to using PWT 6.2, the volatility
coefficient becomes even smaller and quite far from being significant (column 3). The same is true
if we use the original dataset but drop countries with a data quality grade of D (column 4) or if we
drop the same countries while using version 6.2 (column 5).24
Jones and Olken (2005)
This paper, published in the Quarterly Journal of Economics in 2005, relates growth to leadership. Specifically, it estimates the effect of random leader deaths on a country’s growth rate.25 The
two key findings are: such deaths have a significant impact on growth on average; and second, the
death of leaders has a significant impact in autocracies but not in democracies. This paper used
annual growth data from PWT 6.1 and their results are replicated in column 1 of table 8.
When we re-estimated the core specification using data from PWT 6.2 (table 8), we obtained
the following differences compared with the original: First, in the original paper the coefficients of
the random leader death are significant for the year of the leader death and for the two subsequent
years. When we re-estimated it, using new data, the contemporaneous effect remained significant,
but the effect for the two subsequent years ceased to be significant (p-value of 0.12 and 0.13,
respectively); see column 3. Second, the striking differences are related to the disaggregation of
deviation of the residuals of a country across time, and  is a vector of controls, including initial investment share
of GDP, initial population growth rate, initial human capital, initial per capita GDP, lagged GDP, and several time
trend and dummy variables.
24
Dawson et al. (2001), and Katayama and Ponomareva (2009) also show that variability in growth data across
different versions of PWT renders the volatility-growth relationship unstable.
25
The equation estimated is:  =    +     +   +   +  , where  is growth of per capita GDP and
  and   are country fixed effects and time effects, respectively. For each leader death at , there are location-specific
time dummy variables equal to one in one of  − 5  − 4      − 1  + 1  + 2      + 5 and equal to zero otherwise.
These vectors of dummy variables are denoted   and   .
16
the results by type of political regime. In Jones and Olken (2005), random leader deaths had a
significant impact in autocracies in the year of the leader death and in the two subsequent years
(column 1 in the third panel, labeled Autocrats). In the robustness check of column 3 in the same
panel, they were not significant for any of these three time horizons. Even more interestingly, Jones
and Olken find no significant effects in democracies. In the re-estimation using version 6.2 of the
table, leader deaths are significant in the year of the leader death and the following year (column 3
of the fourth panel, labeled Democrats). Thus, not only the average effect but the pattern across
political regimes seems to vary depending on whether data from PWT 6.1 or 6.2 are used.
The same pattern held when we dropped from the sample those countries categorized as very
poor quality (grade D) from a data point of view (column 5 of table 8), i.e., their original results
hold (although the coefficients are smaller) with PWT 6.1 but the reversal of results is still the case
with version 6.2.26
Hausmann, Pritchett, and Rodrik (2005)
This paper, published in the Journal of Economic Growth in 2005, identifies a set of countries
that are deemed to have sustained growth over a long period of time. The paper used data from
PWT 6.1. The criteria used to define sustained growth are that countries must have experienced:
an improvement in growth rates of at least 2 percentage points per capita (this captures the idea
of acceleration); sustained growth of at least 3.5 percent per capita for seven years; and a higher
post-acceleration income level than the pre-acceleration peak (this is to ensure that accelerations
are not simply a rebound from a prior period of very bad performance, for example, due to wars or
conflict or other shocks). In addition, growth per capita must remain above 3 percent after seven
years, which captures the sense that good performance is sustained.
On this basis, Hausmann, Pritchett, and Rodrik (2005) identified 82 country-periods (see their
table 3) that met these criteria.27 Our exercise was simply to see how the list of countries changes
when the data from PWT 6.2 are used; the results are in table 9. Between PWT 6.1 and 6.2, 100
country-periods were identified as sustained growers under this definition. However, only 65 cases
were common to both data sets. In other words, there is a discrepancy of 35 percent between the
two versions. There were 17 cases that Hausmann, Pritchett, and Rodrik (2005), using PWT 6.1
26
In private communication, Jones and Olken showed that their results continue to hold when nonparametric methods are used to estimate their basic specification, or when the original sample is maintained by filling in missing values
for countries for which PWT 6.2 did not report estimates. We focused on their core linear estimation specification.
27
They actually identified 83 growth transitions but one of them has been excluded because data for countryregionplaceBotswana does not go farther back than 1970 in PWT 6.2.
17
(for example, India in 1982 and Sri Lanka in 1979), identified as sustained growers but that did not
show up as such using PWT 6.2. Conversely, there were 18 instances that were identified by PWT
6.2 as sustained growers but that were not so categorized by Hausmann, Pritchett, and Rodrik
(2005) using PWT 6.1. One way of illustrating this discrepancy is to note that only in about 65
percent of the cases was there agreement between the two datasets; in other words, in 4 out of 10
instances, the answer depends on which data set is used (table 9).
Miguel, Satyanath, and Sergenti (2004)
This paper, published in the Journal of Political Economy in 2004, attempts to identify and
quantify the impact of economic factors on civil wars. The estimation is restricted to 41 countries
in Africa and uses an annual panel framework using GDP data from PWT 6.1.
The key results of the paper are that in an ordinary least squares (OLS) or probit framework,
contemporaneous and one-period lagged growth does not have a statistically significant effect on
civil conflict in Africa.28 However, when economic growth is instrumented with rainfall, one-period
lagged growth has a statistically and economically significant effect on conflict–see column 1 (OLS)
and column 4 in their table 9 (IV) from the original.
When we re-ran these regressions using data from PWT 6.2 (our table 10), we found that the IV
results of Miguel, Satyanath, and Sergenti (2004) remained unchanged. However, the OLS results
changed. Specifically in the OLS framework, contemporaneous economic growth, which did not
have a significant effect in the OLS, becomes borderline statistically significant (in column 3, the
coefficient on contemporaneous growth has a t-statistic of 1.67).
Our explanation of these results is that the OLS and probit estimations are more prone to
measurement error (in this case on the right-hand side) which is substantial in the annual panel
framework, especially for Africa. This is cause for substantial differences in results. On the other
hand, a good instrument addresses measurement error, resulting in a more robust estimation.
Aghion, Howitt, and Mayer-Foulkes (2005)
This paper, published in the Quarterly Journal of Economics in 2005, examines the effect of
financial development on income (or technological) convergence. The paper’s two core findings are
that income convergence will depend on the level of financial development; and that there will be a
threshold level of financial development above which countries will converge and below which they
28
The estimated equation is:   =  + 0 + 0  + 1 −1 +   + , where  represents
a vector of controls, and  is instrumented by   and  −1 in the 2SLS specifications.
18
will diverge.
Aghion, Howitt, and Mayer-Foulkes (2004) run a simple cross-country growth regression for a
sample of 63 countries, where the key innovation is an interaction term between the level of financial
development and the initial level of income divergence with the frontier country, the United States.29
One important prediction of the model for which the estimation provides confirmation is that the
coefficient on this interaction is negative (see the second row in column 1 of table 11). When we
re-run the Aghion, Howitt, and Mayer-Foulkes (2004) core regression using PWT 6.2 data–shown
in table 11–we find that while the magnitude of this coefficient drops by about 40 percent, it
remains statistically significant (second row of column 3).
Although the primary prediction of the model is robust to changing data, the second implication
of the Aghion, Howitt, and Mayer-Foulkes (2004) paper is not robust to changing the GDP data.
Aghion, Howitt, and Mayer-Foulkes (2004) derive a threshold value of financial development, which
is the ratio of the coefficient on the income convergence term and the coefficient on the interaction
between income convergence and financial development. In the Aghion, Howitt, and Mayer-Foulkes
(2004) specifications this value is about 25 percent.
In the revised estimations, using PWT 6.2 data, the coefficient on the income convergence term
is economically close to zero and statistically insignificant from zero (in one of the core specifications,
the coefficient on this term switches signs relative to that in the paper). This yields a critical value
for financial development of zero, suggesting that all countries will converge, which runs counter to
the spirit of the main results implied by Aghion, Howitt, and Mayer-Foulkes (2004) (see column 5
of table 11).
Mankiw, Romer, and Weil (1992)
This paper, published in the Quarterly Journal of Economics in 1992, tests an augmented
version of the Solow growth model, with the augmentation consisting of adding human capital
as an additional input (apart from capital and labor) into production. The test comprises two
elements: first, checking for the significance of the coefficients on savings (proxied by investment),
human capital, and population growth (with the latter augmented to reflect technical progress
and adjusted for depreciation); and second, importantly, checking for the magnitudes of these
coefficients. In particular, the Solow model yields the result that the sum of the coefficients on
29
The equation estimated is:  − 1 =  0 +    +   ( − 1 ) +     ( − 1 ) +  , where  − 1 is the gap in
output between country  and the country at the technology frontier,  − 1 is the gap in per capita GDP growth,
and  is the level of financial development.
19
savings, population, and human capital should add to zero; or, in practice, the coefficient on the
savings term adjusted for population growth should be equal in sign and magnitude to the coefficient
on the human capital term also adjusted for population growth.
Mankiw, Romer, and Weil (1992) estimate a simple cross-section regression in which the lefthand side is the level of log per capita PPP GDP.30 They estimate it for three samples, one consisting
of 98 countries for which data are available, another for a sample of 75 countries in which countries
with PWT data quality grade D are dropped, and a third where non-oil countries are dropped from
the sample.
When we re-estimate their equations for the Solow model using PWT 6.2 data (and the non-oil
sample), we find that the coefficients on savings, population growth, and human capital are all
correctly signed and statistically significant (table 12). However, we find that the key test of the
Solow model–that the coefficients on savings and human capital (both adjusted for population)
should be equal in magnitude–comes close to being rejected by the data (in column 5, the p-value
for the test of the Solow restriction is 0.166). This result is interesting because it contrasts with the
general finding that long-horizon cross-sectional growth studies as well as those that use the level
of GDP do not see significant changes in the results. In this study, the results change presumably
because the regressions use the investment-GDP ratio, which is particularly prone to variability.
D. Discussion
Is there an explanation for the pattern observed? The least robust papers are those that use
annual panel data, for which measurement error is large. Conversely, all the papers that survive
the robustness are those that use low frequency data (i.e., cross-sectional estimations) or five-year
panel data. Evidently, averaging reduces “data revision error” and makes the estimations more
robust.
But this raises another question. As long as this “data revision error”–which is in principle a
particular kind of measurement error–is random, why isn’t it captured by the random error term
in the growth regression? The answer is twofold: First, when the PPP-adjusted GDP appears as
the independent variable (in level or growth) in the regression estimation, revision errors to it are
not random and suffer from the biases imposed by the PWT methodology as discussed in section
30
The equation estimated is: ln [ ] = 0 + 1 [ln( ) − ln( +  + )] +  2 [ln() − ln( +  + )] + ,
where   is GDP per worker,  is the investment share of GDP,  is the population growth rate,  is the
technology growth rate,  is the rate of capital depreciation, and  is average number of working age adults in
school.
20
3. For example, whether the countries used in the estimation sample are small or large or whether
the data used extends to the early 1960s biases the PPP-adjusted GDP measures and therefore
cannot be appropriately captured by the random error term. Second, a typical growth regression
usually incorporates regressors that are either directly related to PPP-adjusted GDP (i.e., initial
GDP), or components thereof (such as consumption, investment, or government spending–directly
measured by PWT as well). This introduces measurement error in the dependent variable, which
although challenging could potentially be corrected econometrically. That an error-in-variable is a
probable cause is supported by the fact that the 2SLS results of Miguel, Satyanath, Sergenti (2003)
survive the robustness checks–despite using annual panel data. In Miguel, Satyanath, Sergenti
(2003), the core result is not robust to data revision in the OLS version but it is robust in the IV
version. Having a good instrument evidently helps overcome measurement error.
5
Valuation: Are PWT GDP Estimates Really PPP-Based?
This is an odd question to ask. The raison d’être of the PWT is to come up with PPP prices and
use them for computing GDP estimates. For the benchmark year, PPPs are calculated from the
price data that are collected and these PPPs are used to compute GDP. So, in the benchmark year,
the GDP estimates are at international prices. But matters turn out to be much murkier for years
other than the benchmark year.
Recall equation 3:
d 9596 =  ̂9596 +  ˆ9596 +  ̂9596 

In this equation, the shares of the different components in DA are at international prices and
computed by the PWT. However, the growth rates of the components (C, I, and G) are obtained
from the national income accounts. As such, these growth rates are at domestic, not international,
prices.
It is possible to further decompose equation 3 to see how the additions to domestic absorption
are being valued. In technical appendix 2, we illustrate with a simple example that when there is
more than one good, equation 3 results in valuing the quantity additions to each of the components
of domestic absorption at some hybrid of domestic and international prices (see equation B5 in
technical appendix 2). In general, this will be different from valuing these quantity changes at
international prices and this difference is likely to be systematically larger for smaller countries.
21
This leads to the striking conclusion: Not only are growth rates of GDP not at international
prices, but because of this fact, level GDP estimates are also not fully at international prices for
nonbenchmark years.
Given the documented problems with generating a reasonable PPP-adjusted growth GDP series,
one possibility is for researchers to use the PWT estimates for the levels and national income
estimates for GDP growth rates. In this view, the way forward is to use PPP prices for valuing the
level of GDP and to use domestic prices for valuing growth rates.31
However, there are two problems with this approach. First, if the PWT allows cross-sectional
comparisons at two different points in time based on PPP prices, consistency requires that the
change in these levels across time (the growth rates) should also be at PPP prices. Current estimates
in the PWT do not allow for such consistent intertemporal comparisons because of the valuation
problems we identified.
Second, there is an even deeper problem in the current methodology. Level GDP estimates
in the nonbenchmark years are derived from growth rates that are not at PPP prices. Therefore,
level GDP estimates in such years are themselves not at PPP prices. So, under current PWT
methodology, even pure cross-sectional comparisons are not entirely valid, undermining the basic
rationale of the PWT approach. Note that since GDP levels for nonbenchmark years are calculated
by applying growth rates to benchmark year level GDP estimates, the farther away we move from
the benchmark year the more the level GDP estimates will be composed of quantities that are not
valued at international prices.
6
Thoughts on an Alternative PPP-GDP Chained Growth Series
The existing PWT PPP-GDP series suffer from three shortcomings. First, as documented previously, the methodology leads to large and systematically biased variations across versions of the
table. Second, estimates of PPP-adjusted GDP growth rates and PPP-adjusted level of GDP
for the nonbenchmark years are not at international but rather some “hybrid” prices, which goes
against the raison d’être of PWT. Third, each new generation of PWT leads to discarding useful
information from the disaggregated price data in all previous ICP benchmark studies.
Is there a way of calculating GDP estimates that overcome these problems? In principle, yes.32
31
For example the growth rates in the World Bank’s WDI are all at domestic prices.
Feenstra, Ma, and Rao (2009) have proposed a theoretically sophisticated measure for intertemporal comparisons
but it only applies to consumption and not to the other components of domestic absorption.
32
22
First and foremost, the alternative approach could use the data and the estimates of the level of
GDP compiled in the benchmark years (i.e., 1980, 1985, 1996, and 2005).
The current methodology starts with the level of GDP in the benchmark year, then calculates
the growth rates, which are in turn used to calculate the level of GDP in the benchmark years. We
would propose doing it the other way around: to use the level estimates for the benchmark years
and calculate the growth rates from these level estimates. Take the PPP-adjusted GDPs calculated
for 1985 and 1996 from disaggregated ICP data. While cross-sectional comparisons can be made for
each of the two years, there cannot be intertemporal comparisons based on these estimates because
they are not at common prices. The aim, therefore, is to come up with GDP level estimates that
are at common prices.
In what follows we offer some preliminary thoughts on one possible alternative method in
producing estimates of PPP-adjusted GDP growth rates. We qualify that this is only an attempt to
provide a possible road map rather than to proposed a complete alternative methodology, certainly
beyond the scope of this paper.
Assume that the level of GDP for a country at international prices in the benchmark year 1985
is:
85 = 185  185 + 285  285 + 85  85 
(5)
where the s are all international prices calculated from the Geary-Khamis aggregation, and superscripts refer to the type of consumption good (1 or 2) or to the investment good i (for expositional
simplicity, we assume that GDP comprises only C and I, and that C comprises only 1 , 2 ).
The level of PPP-adjusted GDP in the second benchmark year is:
96 = 196  196 + 296  296 + 96  96 
(6)
To calculate growth rates, we would ideally need to value either period 1985 quantities at 1996
prices or 1996 quantities at 1985 prices. Thus we can derive two new GDP estimates that are,
respectively:
2
85
= 185  196 + 285  296 + 85  96 
(7)
1
96
= 196  185 + 296  285 + 96  85 
(8)
The point is that the data for computing Y according to equations 7 and 8 are generated by
the PWT. For each benchmark year, the quantities of the different consumption and investment
23
goods are available at the disaggregated level (recall that these are the inputs for the Geary-Khamis
aggregation procedure, and these quantities are derived simply by deflating expenditures on each
good at domestic prices by domestic prices). And for each benchmark year, the Geary-Khamis
procedure yields disaggregated international prices.
Equations 5 and 8 will yield a GDP series that is intertemporally comparable (at period 1985
prices); equations 6 and 7 will also yield a GDP series that is internationally comparable but at
period 1996 prices. It is now simple to calculate growth rates.
We can then derive the growth rates of GDP from 1985 to 1996 either at base-year prices, which
is:
̂85 =
∆1  185 + ∆2  285 + ∆ 85

85
(9)
∆1  196 + ∆2  296 + ∆ 96

2
85
(10)
Or at current year prices which is:
̂96 =
Our preferred approach, however, would be to calculate a chained series that can be used for
international comparisons. This would involve averaging the two growth rates emerging from
equations 9—10 to yield a chained growth estimate between these two years. The way we would
envisage is that the chaining would be based on successive ICPs. Thus, there would be a growth
estimate between 1985 and 1996 based on the ICPs in 1985 and 1996. Similarly, there would be an
estimate between 1996 and 2005 based on the ICPs in these two years. And so on. The advantage
of this approach would be that every time a new ICP is conducted, it would not change historical
growth rates of GDP. Only growth rates going forward until the next ICP would be estimated.
Of course, a number of practical complications will have to be addressed. First, the basket of
disaggregated goods varies across benchmark years, so some way of deriving a common basket will
have to be found (i.e., equations 5—8 will need to have a common basket of goods). In other words,
we would need the detailed price data for the composites of C, I, and G aggregates between the
two benchmark years; although such data are available for OECD countries, they are not readily
available for non-OECD countries.
A second problem with our proposal is that it would only provide a growth rate for the horizon
between successive ICPs (either 10 years or over 5 years, if in the future ICPs are conducted every
5 years). There will not be a high frequency (i.e., annual) growth rate. But in assessing whether
this is a major loss, two points should be kept in mind: annual data seem particularly problematic
24
because of the current extrapolation methodology. Moreover, the motivation of the PWT is to make
meaningful comparisons of standards of living across countries at a given point in time. Taking
long averages seems to preserve this scope but high frequency data (especially annual frequency
comparisons) become unusable.
A third problem relates to samples. For 1985 and 1996, but especially for previous benchmark
years, the sample of countries for which ICPs have been done remains limited. How to derive the
chained growth numbers for a large enough sample of countries going back before 1985 or 1980
needs further thought.
Once we have a chained growth estimate of the type we have proposed, this is how we would
envisage the use of different PWT GDP series. First, for pure cross-country comparisons of the
level of per capita income, researchers should use the estimates from the different benchmark years.
Thus, if a researcher wants to compare incomes across countries in 2005, the benchmark estimate
for 2005 should be used. If the time period is 1996, the benchmark estimate from the ICP estimate
of PWT 6.1 (not 6.2) should be used. And if the time period is 1985, the estimate from PWT 5.6
for 1985 should be used. The key message here is that researchers should use that version of the
PWT that is closest to the timing of inquiry and should not use the most recent version. That is,
if the year of inquiry is 1985, the most recent PWT series (PWT 6.1 or 6.2) should NOT be used
because the level estimate for 1985 in PWT 6.1 and 6.2 is NOT at international prices. Of course,
if the year of inquiry is say 1990, researchers could use either the PWT 6.1 estimate for 1996 or
the PWT 5.6 estimate for 1985.
Second, if the object is intertemporal growth comparisons, especially growth over the medium
and long run, researchers should use the new chained growth estimates that we have proposed.33
Our third suggestion relates to the use of annual data. Given that researchers will continue to
use annual growth data, should they use the PWT series or some alternative series such as that
from the WDI? It must be clearly understood that annual growth series in both PWT and WDI
are not at PPP prices. In the WDI, all growth (even the PPP series) is measured at domestic
prices and not at PPP prices. In the PWT, annual growth is measured at hybrid (domestic and
33
Of course, these chained growth estimates could, in principle, lead to a level GDP series (at least for the benchmark
years) that is at a common international price. Thus, based on a chained growth estimate based on 1985 and 1996
ICPs, we could have a chained level estimate for 1985 that applies the new growth estimate to the level estimate in
1996, and extrapolates backwards. Or, we could have a new chained level estimate for 1996 that applies the new
growth estimate to the level GDP numbers for 1985. But for pure cross-country comparisons, we would urge that
researchers not use the level estimates derived from this chained growth procedure. Rather, they should follow the
suggestion earlier about using that version of the PWT that is closest to the period of interest.
25
PPP) prices. But the deficiencies we have identified in the PWT estimates are not overcome by
using WDI data. Given the sensitivity of the growth results to data revisions at annual frequency,
we would emphasize that research using annual data should demonstrate robustness to alternative
data series. That would be our preferred research strategy. Alternatively, if we had to choose one
series for research based on annual data, we would probably favor GDP growth data (constant
price series either in local currencies or dollars) from the World Development Indicators for two
reasons: first, at annual frequency, PPP effects are less important so the costs of foregoing the
use of PPP-adjusted data are smaller. Second, WDI growth rates are subject to only one layer
of revision, namely by the country authorities, whereas, the PWT further revises these national
growth rates before they become inputs for the PPP-based growth estimates. If the PWT growth
estimates were truly at PPP prices the benefits may outweigh the costs of additional layers of
revisions but that is not the case as we have shown. But the use of WDI data must come with the
clear recognition that the data are not strictly speaking comparable across countries because they
are not at common international prices.
7
Conclusion
The PWT suffers from problems of variability and valuation. There is considerable variability in
the level and growth of PPP-adjusted GDP estimates and in the estimates of the PPPs across
alternative versions of the PWT. This variability stems in part from changes to underlying national
income accounts data but is also systematically related (inversely) to the size of a country and to
the distance of the data from the benchmark data. Because this variability is intrinsic to the PWT
methodology there is little basis for knowing whether newer versions are better than older versions.
In fact, the distance from benchmark finding suggests that in every new version, historic data tends
to become more variable.
This variability is such that some standard results in the growth literature are not robust across
alternative versions of the PWT. It does appear safe to use PPP GDP level data, when looking
at cross-sections. Long-run changes, e.g., over 30 years, also appear to be robust. Medium-run
growth rates, such as 10-year panels, may also be safe, although we flag this as a point for further
investigation.
In terms of robustness to data revisions across PWT version, it is generally not safe to use
annual data. The exception would be for countries with quality grades of A and B or if there
26
is a good instrumentation strategy. In general, annual data from non-OECD countries should be
treated with caution.
Thus, growth studies should demonstrate robustness to different versions of the PWT, especially
where high frequency data are used. Where possible, robustness should also be demonstrated for
samples of countries for which benchmark data are available and for samples excluding the small
countries or those with a quality grading of C or D. If results do not survive one or both of these
checks, a much bigger health warning should be attached to any policy implications.
On valuation, we also found, surprisingly, that, for years other than the benchmark year, GDP
growth and level estimates from the PWT are not at PPP prices. To overcome this problem as
well as the systematic variability (with respect to country size and distance from the benchmark
year) of the GDP numbers, we suggested an alternative approach to calculating a chained growth
estimate.
It remains to be seen how version 7.0 of the PWT will confirm or change any of this assessment.
It is also unclear whether that version will definitively supersede all previous versions (and all
alternatives, including the WDI).
27
Technical Appendix 1
The Penn World Table Calculation of the Level and Growth of PPP-Adjusted GDP
We outline below the construction of the purchase power parity (PPP)-adjusted level of GDP
and its growth rate. In the process, we highlight some key and relatively unknown facts about
PWT data construction. The description below refers to version 6.1 of PWT.34
The procedure used by PWT to construct PPP-adjusted GDPs can be stripped down to two
main steps. In the first step, international price levels and PPP-adjusted GDPs are calculated for a
benchmark year (for PWT 6.1 in 1996). The second step relates to calculations for nonbenchmark
years. To do this, international prices for the benchmark year are extrapolated backward and
forward to produce PPP time series for each country considered. Then PPP-adjusted GDP levels
and growth rates are calculated for all countries and years using the relevant PPPs from step 2.
Next, we look into each of the steps, focusing on what is essential in the final construction of the
PWT series.
Step 1: Obtaining PPPs and PPP-adjusted GDPs for the benchmark year
All calculations start for a benchmark year (the year for which detailed price data are collected;
for PWT 6.1 in 1996) and for benchmark countries (the countries that participate in the International Comparison Program [ICP]); that is, benchmark countries are those for which actual and
detailed price data are collected.35
The raw data for calculating PPPs for the benchmark year are:
1. Expenditures in local currencies () (obtained from the national income accounts), where
i refers to categories of goods and services and j to the country; in PWT 6.1 there were 31
basic categories of goods and services.
2. The individual prices  for these basic goods and services categories in each country covered
in the ICP exercise.
Dividing these expenditures by the individual prices yield notional quantities for each of the
categories (i.e.,  = ()  ).
These data are used to derive two sets of international prices:
1. A set of international prices for each of the 31 categories of goods and services (  ); that is,
these prices are the same for each category for all countries.36
2. Purchasing power parity for each country ( ); PPPs for each country.
34
Special procedures are used for certain countries (i.e., China and India) at various stages. These are not described
here and the interested reader is advised to consult the technical appendix to the PWT 6.1.
35
See table 1 for a list of benchmark years and benchmark countries used for different versions of PWT.
36
International prices for each category are expressed relative to the United States;   = 1 for every .
28
How are the (  )s and ( )s derived? They are obtained from a system of two sets of simultaneous equations, which is also known as the Geary (1958) and Khamis (1972) (GK thereafter)
aggregation procedure. These equations follow from the definition of the two sets of prices as
follows:
Ã
"
!#
X µ  ¶

P

(A1)
 =

 

Equation A1 defines the international price for each commodity. The first expression on the
right-hand side (  ) is the price of a category of goods i in country j deflated by that country’s
P
PPP, and the second (    ) is the share of i in country j in the world consumption of that
good. The equation makes clear that the international price for each of the 31 commodities is the
weighted average of their domestic prices, where the weights are the share of each country in total
world consumption of that commodity.
P
()
 = P

(A2)
   
Equation A2 defines the price level of a country or its PPP. These relative country prices are
the heart of the PWT, subsequently used to produce cross-country comparable GDP measures. For
each country, the total expenditure in domestic prices divided by the sum of expenditures on each
commodity at international prices yields the PPP for that country.3738
As a result of the GK procedure, in PWT 6.1 there are international prices for 31 individual
commodities, denoted as (  )s, and a set of country-specific PPPs, denoted as ( )s.
A final point is worth making here. The 31 categories of goods and services are also placed into
three more aggregate categories corresponding to the three basic categories of domestic absorption
(DA)–consumption (C ), investment (I ), and government (G). For each country j, the countryspecific PPPs for C, I, and G (     ) are also calculated. These will vary across countries
simply because the composition of the 31 categories of goods and services (for which there are
common international prices) will vary across countries.
These prices are used to calculate the PPP-adjusted GDPs for the benchmark year. GDP is
calculated as the sum of real domestic absorption (measured at international prices) plus the net
foreign balance. The constituent elements of domestic absorption–C, I, and G–are also measured
37
Once the PPPs are calculated for the benchmark countries for the benchmark year, the PWT proceeds to calculate
the same for non-benchmark countries. This is done essentially through a two-step regression procedure, which yields
coefficients that can be used to estimate the PPPs for non-benchmark countries. For the purpose of this paper, the
estimation procedure is not relevant and is therefore skipped.
38
There are also historical data on PPPs obtained from previous ICP rounds. For example, while for PWT 6.1 the
ICP was done for 115 countries for the year 1996, a similar exercise was done in 1985 for 64 countries for version PWT
5.6. These 1985 PPPs can be extrapolated using national income accounts deflators for consumption, investment,
and government to yield PPPs for 1996. So, for a number of countries, there are multiple data on PPPs for the
benchmark year (1996). The PWT then utilizes a weighting method for these multiple data sources to arrive at a
final set of international prices or PPPs for the year 1996 for all countries (see pages 9—11 of technical appendix to
the PWT 6.1 for details).
29
at international prices by multiplying the quantities of C, I, and G by their respective international
price obtained above.
(A3)
 =    
where  is consumption at international prices,   is international price of consumption and 
is the notional quantity of consumption. By a similar procedure real I and real G are produced.
Summing these three yields real domestic absorption as follows:
 =  +  +  
(A4)
 =  +    
(A5)
 is domestic absorption in international prices for country j, Y is PPP-adjusted GDP and NFB
is net foreign balance. Note that the net foreign balance does not require the calculation of PPPs
as they are values at the price level of DA (see footnote 8).
Step 2A: Calculating international prices for the non-benchmark years
Once PPPs for the benchmark year are calculated, the PWT extrapolates these PPPs going
forward and backward in time. How does it do so? Take PWT 6.1 and the year 1995, which is one
year before the benchmark year. For 1995, the PWT obtains nominal expenditures for the three
components of domestic absorption–consumption, investment, and government expenditure–from
the national income accounts. Each country j will also have a price level associated with C, I, and G
(     , respectively) for 1995. These are just the 1996 PPPs for each of these categories,
calculated as shown above, deflated by the price change between 1995 and 1996 for C, I, and G,
where the price changes are obtained from national income accounts.
These extrapolated domestic prices for the three components of domestic absorption obtained
(  ) are used in place of the detailed price data collected for the benchmark year. However, unlike in the case of the benchmark year where detailed prices are collected for a large set of
commodities, for the nonbenchmark years, each country has only three international prices.
With these prices and with national income data for 1995 on consumption, investment, and
government expenditures, the PWT uses the GK aggregation procedure to calculate international
prices ()s for C, I, and G, as well as the PPPs ()s for all countries exactly as in step 1.
These international prices are then used to convert consumption, investment, and government
expenditures at domestic prices into expenditures at international prices as in equation A3 in step
1.
Step 2B: Calculating the level of PPP-adjusted GDP and its growth rate for nonbenchmark years
Step 2A yields real C, real I, and real G for all countries for nonbenchmark years. How is the
level of PPP-adjusted GDP then calculated, say for 1995? Essentially via a circuitous procedure
that first calculates the growth rate of domestic absorption between 1995 and 1996 at international
prices and then applies this growth rate to the 1996 level of DA to derive the level of DA in 1995.
30
To this, the net foreign balance for 1995 is added to obtain the level of PPP-adjusted GDP for
1995.
d 9596 =  ̂9596 +  ˆ9596 +  ̂9596 
(A6)

´
³
d 9596 
(A7)
95 = 96  1 + 
95 = 95 +  95 
(A8)
where DA is domestic absorption in international prices, the hat sign over a variable denotes growth,
 ,  , and  are the shares of consumption, investment, and government spending, respectively,
in domestic absorption, and NFB is net foreign balance.
Three points are worth emphasizing about equation A6. First, the growth rates of real C, I,
and G, are from the national income accounts and mostly do not change across PWT versions.
Because they are from national income accounts, these growth rates are at domestic, not PPP,
prices. However, the weights assigned to each of these components, the shares of each component
in domestic absorption, are measured at international prices in 1995 which are obtained from steps
1 and 2. For example the share of C is given by:
 =
  

=

 + +
(A9)
Second, the PWT computes two PPP-adjusted GDP series,39 the chained series (RGDPCH), which
is the most commonly used and the one recommended by the authors of the PWT, and an RGDPL
(or Laspeyres series). The difference between the two is simply that in the RGDPCH, the shares
change for every year the growth rate is calculated; hence note that the shares are time sensitive.
In contrast, in the RGDPL series, the shares remain the same for all years and are the shares
(calculated at international prices) for the benchmark year. Therefore, when the PWT is revised,
the shares are revised for the benchmark year and all subsequent years. But for the RGDPL series,
the changes in nonbenchmark years are not relevant.
Third, this leads to understanding why the growth rates differ between the two series in the
PWT and in turn how these differ from the PPP-adjusted growth rate calculation in the WDI. In
the WDI, the growth rate will typically NOT change across revisions (other than to reflect revisions
of national accounts) because the growth rates are calculated from the national income accounts.
In terms of equation A4, PWT and WDI use the same numbers for real consumption, investment,
and government growth, all obtained from the national income accounts. It is the shares that
are different: in the WDI, these shares are from the national income accounts themselves, and
therefore measured at domestic prices, and change with time; for the RGDPL series, the shares are
at international prices but fixed at the levels of the benchmark year; for the RGDPCH series, the
shares are at international prices and change every year based on changing international prices.
Steps 2A—2B are then repeated for each of the years before and after the benchmark year to
yield PPP-adjusted GDP growth rates and levels for these years.
39
There is a third series called RGDPTT that is not discussed here.
31
Technical Appendix 2
Valuation issues in the Measurement of GDP Growth in the PWT
How are growth rates computed in the PWT? To answer this question and focus on the key
points, we assume that GDP comprises consumption (C ) and investment (I ) and is hence equal to
domestic absorption (DA).
One consumption and one investment good
We start with the case where there is only one consumption and one investment good. First,
some notation. The quantities of these goods are denoted by small case letters c and i, respectively.
Subscripts, which apply only to value and price variables, refer to whether they are measured at
domestic or international (i.e., purchasing power parity) prices. Thus P dom refers to domestic price
and P int to international prices. Superscripts refer to the year of measurement, and can either
be the current year (T ) or the base year (B). For example,   denotes real consumption for the
year T. For the price variables, there will be two superscripts, the first referring to the year of
 refers to
measurement and the second to the good (consumption, c, or investment, i). Thus 
the price of the consumption good measured at domestic prices for the base year. Thus,
 =  =  +  = 
̂  =
 ∆ 
 ∆ 




+

 
 





(B1)
(B2)
where the hat sign over a variable denotes growth. Equation B2 simply says that GDP growth in
year T is a weighted average of growth of consumption and investment. In the PWT’s chain series
(RGDPCH), the weights are the shares of C and I in domestic absorption measured at current
international prices. So, if the growth rate is calculated for the period 1993—1994, the weights are
at 1993 PPP prices obtained from the Geary-Khamis aggregation procedure described in step 1 in
∆ 
∆ 
technical appendix 1. Note that the growth rates of C (   ) and I (   ) are obtained from


the national income accounts and are therefore at domestic base-year prices.
Equation B2 can be rewritten as:
̂  =
  ∆  
  ∆  
 
 


+





 
 


which in turn simplifies to:
̂

¢
  + ∆   
∆ 

=



¡
(B3)
Equation B3 shows that the PWT chain series for the growth rate essentially involves valuing
the additions to (the quantities) of consumption and investment at current year international prices.
This is a chain-weighted index because the prices used for valuing these additions change every year.
As a result, the RGDPCH series does not use benchmark year international prices; it uses current
year international prices that are obtained in the PWT by extrapolating from the benchmark years
32
(see technical appendix 1, step 2) and applying the GK aggregation procedure. These prices change
with every revision as shown below and the revisions are not random but systematically related to
country attributes.
A second feature of these current year international prices is that they are calculated at a
highly aggregated level, namely at the level of aggregate consumption and investment. For the
benchmark year, aggregate consumption is obtained by adding up consumption of the different
goods for which disaggregated price data are available. For nonbenchmark years, it is as if there is
only one consumption and one investment good.
Two consumption and one investment good
Does this aggregation affect the calculation of growth rates? Suppose there are two consumption goods and one investment good. The spirit of the PWT suggests that all three goods should
be valued at international prices. But how are they actually valued? If we had disaggregated international prices, we can write down how growth of DA should be computed. Essentially, equation
B2 should be rewritten to take account of the extra consumption good. Thus,
¡ 1 1
¢
 2 + ∆   
∆  + ∆ 2 



(B4)
̂ =


Now, the second superscript is 1 or 2 for the two consumption goods and i for the investment good.
Equation B4 is just an extension of equation B3 and says that GDP growth is obtained by valuing
each of the quantity changes (to the consumption goods and investment good) at their respective
current international prices.
But for nonbenchmark years, we do not have disaggregated international price data. GDP
growth (̂2 ) is measured as follows:
µ
 1 + ∆  2 ¶

 ∆ 
∆
̄





+
̂2 = 

 


̄

which can be simplified to:
(¡
¡ 1
¢
¢)
 2 ̄  
 1  1 + ∆ 2  2


∆
+

∆ 




̂2 =


+


  

( 1 +  2 ) ̄


and further to:
̂2
1
= 

½
µ 1
¶
µ 2
¶¾

∆ 
 1   
 2   
∆
+
∆
+
̄
̄







̄
̄
(B5)
Equation B5 shows that each of the quantity changes is valued not at current year international
prices (as in equation B4) but at some hybrid of domestic and international prices. The bar sign
over a variable denotes that it refers to the composite good as does
´o superscript C. The
n³the1second

which is the domestic
price term in brackets for the two consumption goods consists of

̄
relative price of that good (that is, it the domestic price of good 1 relative to the average price of
33
  is the average international price of the composite of
the composite of goods 1 and 2); and ̄
goods 1 and 2.
To more clearly identify the difference between how GDP growth ought to be measured and
how it is, we can take the difference between equations B4—B5, which yields:
̂  − ̂2 =

̄


½
µ 1
µ 2
1 ¶
2 ¶¾




2
∆ 1
+
∆

−
−




̄
̄
̄
̄
(B6)
This equation shows that the difference between the two depends on how different the relative price
 1
is of a consumption good at domestic prices ( ̄
 ) from its relative price at international prices

 1
( ̄ 
  ).


Note that this difference will vary across time because domestic prices are computed for a
fixed base period (which can be different across countries) while the international prices are current
prices.
We know that this difference in the relative prices will vary systematically across countries.
It will be greater for smaller countries because under the GK procedure domestic prices of larger
countries have a greater weight when computing international prices. This is the Gerschenkron
effect. Thus, GDP growth rates are likely to be measured with greater error (relative to the true
growth rate that is consistent with the spirit of the PWT as represented in equation B4) for smaller
countries.
A second problem is that the farther away T is from the base year, the greater the discrepancy.
Hence growth rate calculations for years farther away from the benchmark year are likely to have
greater measurement error.
34
References
Acemoglu, Daron, Simon Johnson, James Robinson, and Yunyong Thaicharoen. 2003. “Institutional Causes, Macroeconomic Symptoms: Volatility, Crises and Growth.” Journal of
Monetary Economics 50: 49—123.
Aghion, Philippe, Peter Howitt, and David Mayer-Foulkes. 2005. “The Effect of Financial Development on Convergence: Theory and Evidence.” Quarterly Journal of Economics 120:
173—222.
Balassa, Béla. 1964. “The Purchasing Power Parity Doctrine: A Reappraisal.” Journal of Political
Economy 72: 584—596.
Barro, Robert J. 1999. “Determinants of Democracy.” Journal of Political Economy 107: S158—
S183.
Burnside, Craig, and David Dollar. 2000. “Aid, Policies, and Growth.” American Economic
Review 90: 847—868.
Ciccone, Antonio, and Marek Jarocinski. 2010. “Determinants of Economic Growth: Will Data
Tell?” American Economic Journal: Macroeconomics, forthcoming.
Dawson, John W., Joseph P. DeJuan, John J. Seater, and E. Frank Stephenson. 2001. “Economic Information Versus Quality Variation in Cross-Country Data.” Canadian Journal of
Economics 34: 988—1009.
Deaton, Angus. 2006. “Purchasing Power Parity Exchange Rates for the Poor: Using Household
Surveys to Construct PPPs.” Princeton University Research Program in Developing Studies
working paper. Princeton, NJ: Princeton University.
Deaton, Angus, and Alan Heston 2010. “Understanding PPPs and PPP-Based National Accounts.” American Economic Journal — Macroeconomics, forthcoming.
De Long, J. Bradford, and Lawrence H. Summers. 1991. “Equipment Investment and Economic
Growth.” Quarterly Journal of Economics 106: 445—502.
Demirguc-Kunt, Asli, Luc Laeven, and Ross Levine. 2004. “Market Structure, Institutions, and
the Cost of Financial Intermediation.” Journal of Money Credit and Banking 36: 593—622.
Dowrick, Steve. 2005. The Penn World Table: A Review. Australian Economic Review 38:
223—228.
Dowrick, Steve, and John Quiggin. 1997. “True Measures of GDP and Convergence.” American
Economic Review 87: 41—64.
Easterly, William, Michael Kremer, Lant Pritchett, and Lawrence H. Summers. 1993. “Good
Policy or Good Luck?: Country Growth Performance and Temporary Shocks.” Journal of
Monetary Economics 32: 459—483.
Feenstra, Robert, Alan Heston, Marcel P. Timmer, and Haiyan Deng. 2009. “Estimating Real
Production and Expenditures Across Nations: A Proposal for Improving the Penn World
Tables.” Review of Economics and Statistics 91: 201—212.
Feenstra, Robert, Hong Ma and D.S. Prasada Rao. 2010. “Consistent Comparisons of Real
Incomes across Space and Time.” Macroeconomic Dynamics, forthcoming.
35
Geary, Roy C. 1958. “A Note On the Comparison of Exchange Rates and Purchasing Power
Between Countries.” Journal of the Royal Statistical Society, Series A (General) 121: 97—99.
Gershenkron, Alexander. 1947. “The Soviet Indices of Industrial Production.” Review of Economic Studies 29: 217—26.
Handbook of the International Comparison Program, http://unstats.un.org/unsd/methods/icp/ipc6_htm.
Hausmann, Ricardo, Lant Pritchett, and Dani Rodrik. 2005. “Growth Accelerations.” Journal of
Economic Growth 10: 303—29.
Henderson, J. Vernon, Adam Storeygard, and David N. Weil. 2009. “Measuring Economic Growth
from Outer Space.” NBER Working Papers No. 15199. Cambridge, MA: National Bureau of
Economic Research.
Heston, Alan. 1994. “National Accounts: A Brief Review of Some Problems in Using National
Accounts Data in Level of Output Comparisons and Growth Studies.” Journal of Development
Economics 44: 29—52.
Heston, Alan, Robert Summers, and Bettina Aten. 2002. Penn World Table Version 6.1. Center
for International Comparisons at the University of Pennsylvania (CICUP). October.
Heston, Alan, Robert Summers, and Bettina Aten. 2006. Penn World Table Version 6.2. Center
for International Comparisons of Production, Income and Prices at the University of Pennsylvania. September.
IMF (International Monetary Fund). 2008. World Economic Outlook. Washington.
Jones, Ben, and Benjamin Olken. 2005. “Do Leaders Matter? National Leadership and Growth
Since World War II.” Quarterly Journal of Economics 120: 835—864.
Katayama, Hajime, and Natalia Ponomareva. 2010. “Does the Version of the Penn World Tables
Matter? An Analysis of the Relationship between Growth and Volatility.” Canadian Journal
of Economics, 43(1): 152-79 .
Khamis, Salem H. 1972. “A New System of Index Numbers for National and International Purposes.” Journal of the Royal Statistical Society, Series A (General) 135: 96—121.
Kravis, Irving B., Alan Heston, and Robert Summers. 1978. Real GDP Per Capita for More Than
One Hundred Countries. Economic Journal 88: 215—242.
Levine, Ross, Norman Loayza, and Thorsten Beck. 2000. “Financial intermediation and growth:
Causality and Causes.” Journal of Monetary Economics 46: 31—77.
Mankiw, N. Gregory, David Romer, and David N. Weil. 1992. “A Contribution to the Empirics
of Economic Growth.” Quarterly Journal of Economics 107: 407—437.
Miguel, Edward, Shanker Satyanath, and Ernest Sergenti. 2004. “Economic Shocks and Civil
Conflict: An Instrumental Variables Approach.” Journal of Political Economy 112: 725—753.
Neary, J. Peter. 2004. “Rationalizing the Penn World Table: True Multilateral Indices for International Comparisons of Real Income.” American Economic Review 94: 1411—1428.
Nuxoll, Daniel A. 1994. “Differences in Relative Prices and International Differences in Growth
Rates.” American Economic Review 84: 1423—1436.
36
Ramey, Garey, and Valerie A. Ramey. 1995. “Cross-Country Evidence on the Link between
Volatility and Growth.” American Economic Review 85: 1138—1151.
Rao, Prasada D.S., Alicia N. Rambaldi, and Howard E. Doran. 2009. “An Econometric Approach
to Construct World Tables of Purchasing Power Parities and Real Incomes: Analytical Properties and Tables for 1970—2005,” working paper, University of Queensland.
Rao, Prasada D.S. , and E.A. Selvanathan. 1992. “Computation of Standard Errors for GearyKhamis Parities and International Prices: A Stochastic Approach.” Journal of Business and
Economics Statistics 10: 109—115.
Sachs, Jeffrey D., and Andrew M. Warner. 1995. “Economic Reform and the Process of Global
Integration.” Brookings Papers on Economic Activity 1995: 1—118.
Samuelson, Paul A. 1994. “Facets of Balassa-Samuelson Thirty Years Later.” Review of International Economics 3: 201—226.
Srinivasan, T. N. 1994. “Data Base for Development Analysis: An Overview.” Journal of Development Economics 44: 3—27.
Summers, Robert, and Alan Heston. 1980. “A New Set of International Comparisons of Real
Product and Price Levels Estimates For 130 Countries, 1950—1985.” Review of Income and
Wealth 26: 19—66.
Summers, Robert, and Alan Heston (1984) “Improved International Comparisons of Real Product
and its Composition, 1950-80,” Review of Income and Wealth 30 : 207—261.
Summers, Robert, and Alan Heston. 1991. “The Penn World Table (Mark 5): An Expanded Set
of International Comparisons, 1950-1988.” Quarterly Journal of Economics 106: 327—368.
Summers, Robert, and Alan Heston. 1996. “International Price and Quantity Comparisons:
Potentials and Pitfalls.” American Economic Review 86: 20—24.
Samuelson Paul. 1964. “Theoretical Notes on Trade Problems.” Review of Economics and Statistics 23: 145—154.
van Veelen, Matthijs. 2002. “An Impossibility Theorem Concerning Multilateral International
Comparisons of Volumes.” Econometrica 70: 369—75.
World Bank. 2007. World Development Indicators. Available at the World Bank’s website,
www.worldbank.org.
World Bank. 2008a. Global Purchasing Power Parities and Real Expenditures: 2005 International
Comparison Program. Available at the World Bank’s website, www.worldbank.org.
World Bank. 2008b. ICP 2003-2006 Handbook. Available at www.worldbank.org.
Young, Alwyn. 2009. “Real Consumption Measures for the Poorer Regions of the World.” London
School of Economics working paper. London: London School of Economics.
37
Table 1. Average growth rates in African countries (1975–1999)
Top 10 countries
PWT 6.2
PWT 6.1
Country
Growth
Country
Growth
Countries not appearing on both lists
Botswana
Equ.Guinea
Cape Verde
Egypt
Mauritius
Lesotho
Tunisia
Mali
Ethiopia
Morocco
4.50%
4.00%
3.70%
3.70%
3.70%
3.50%
2.70%
2.00%
1.60%
1.60%
PWT 6.2
Country
Gabon
Zambia
Madagascar
Togo
Guinea-Bissau
Comoros
Niger
Nigeria
Chad
Mozambique
Growth
-2.60%
-2.10%
-1.90%
-1.70%
-1.40%
-1.20%
-0.70%
-0.50%
-0.50%
-0.40%
Botswana
Cape Verde
Mauritius
Egypt
Tunisia
Uganda
Morocco
Lesotho
Congo, Rep of
Malawi
5.10%
4.70%
4.30%
3.70%
2.50%
1.70%
1.70%
1.50%
1.50%
1.20%
Congo, Republic of
Equatorial Guinea
Ethiopia
Malawi
Mali
Uganda
Bottom 10 countries
PWT 6.1
Country
Growth
Countries not appearing on both lists
Equ. Guinea
Mozambique
Zambia
Comoros
Madagascar
Cote d'Ivoire
Niger
Mauritania
Togo
Namibia
-2.70%
-2.40%
-1.80%
-1.60%
-1.40%
-1.40%
-1.30%
-1.30%
-1.00%
-0.90%
Chad
Cote d'Ivoire
Gabon
Guinea-Bissau
Mauritania
Namibia
Nigeria
Togo
Country switching lists:
Equ. Guinea
Notes: This table presents the countries with the top 10 and bottom 10 average growth rates between 1975 and 1999
calculated using PWT 6.2 and PWT 6.1 GDP data (RGDPCH). 40 countries have complete data from 1975-1999.
Six countries appear on the top 10 list for one dataset but not the other. Eight countries appear on the Bottom 10 list
for one dataset but not the other. One country, Equatorial Guinea, switches lists. According to PWT 6.1, it is the
worst-performing country; according to PWT 6.2, it is the second highest performing country.
38
Table 2. The evolution of the Penn World Table (PWT)
PWT version
1
2
3
4
5
5.6
6.1
6.2
7
Benchmark year
1970
1975
1980
1985
1985
1996
1996
2005
Year released
1980
1984
1988
1991
1995
2002
2006
TBD
ICP phases
Phase 1
Phase 2
Phase 3
Phase 4
Phase 5
Phase 6
Phase 7
Year(s)
1970
1973
1975
1980
1985
1996
2005
Number of countries
10
16
35
61
62
115
146
Countries in the PWT
119
115
128
134
151
168
188
TBD
Notes: Penn World Table version 2 was never released. There are some discrepancies in
these numbers across different PWT sources. This table uses information from the datasets
themselves and the PWT 6.1 appendix. Other information can be found in the various
papers published along with releases of versions of the Penn World Table.
39
Table 3. Explaining Revisions to Price Levels Across Versions
Dependent variable
Abs(difference in log P)
[1]
[2]
[3]
Estimation
Dataset Comparison
Sample
Abs(Difference in NIA GDP Growth)
Abs(Difference in PWT-NIA
GDP Growth)
PWT Grade (D=1…A=4)
Total ICP Studies
Log(GDP)
Distance from Benchmark
Year, in Decades (1996)
Distance * Total ICPs
Constant
N
R-squared
[4]
[5]
OLS
PWT6.2 vs. PWT6.1
1970 to 1999
[6]
0.433**
[1.98]
[7]
[8]
0.303
[1.42]
0.0383
[0.169]
0.486***
[7.581]
0.799***
[6.87]
-5.657***
[-15.0]
-2.409***
[-13.0]
0.208
[1.36]
-2.091***
-0.863***
[-10.1]
[-6.19]
5.579*** 6.158***
[11.3]
[10.8]
-1.086***
[-7.69]
8.651*** 8.073*** 23.46*** 16.11*** 61.86*** 3.925*** 25.75***
[35.2]
[22.2]
[20.5]
[22.6]
[11.7]
[9.92]
[7.65]
2845
3016
3016
3016
3016
3016
2845
0.001
0.057
0.085
0.062
0.058
0.062
0.117
0.263*
[1.719]
-0.579***
[-4.066]
5.882***
[10.15]
-1.041***
[-7.254]
17.61***
[5.092]
2845
0.147
***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust t-statistics in brackets.
Notes: The dependent variable is the absolute value of the difference in the log-level of prices (P) across PWT 6.2 and PWT 6.1. NIA
GDP is the GDP per capita, constant price, series from the World Development Indicators dataset (WDI), as reported by national
authorities. WDI 2002 data corresponds to PWT 6.1 and WDI 2005 corresponds to PWT 6.2. PWT-NIA GDP is a GDP per-capita,
constant price, series constructed using the Penn World Table's national income accounts datafiles. The variable PWT grade can take
on four values, 1-4, with 1 representing the worst PWT rating of D, and 4 representing the best PWT rating of A. Total ICP studies
refers to the number of ICP studies a country has participated in. GDP is total GDP and is measured at purchasing power parity. The
distance variable is calculated as the absolute difference between the year of the observation and 1996 (i.e., abs(t-1996)), and is
divided by 10 to express the variable in decades. 1996 is the "Benchmark Year" used in constructing both PWT 6.2 and PWT 6.1. The
sample consists of the 104 countries in the “Long Run Sample” used in other tables and figures.
40
Table 4. Explaining Revisions to Levels of GDP Per Capita Across Versions
Dependent variable
Abs(difference in log per capita PPP-adjusted GDP)
[1]
[2]
[3]
[7]
[8]
0.891***
[3.09]
0.563*
[1.765]
0.603***
[6.017]
-0.838***
[-4.12]
-1.758***
-0.653***
[-13.6]
[-5.06]
2.456*** 3.639***
[6.99]
[5.62]
-0.741***
[-4.66]
11.39*** 10.21*** 21.69*** 17.90*** 55.67*** 9.495*** 27.76***
[45.7]
[33.5]
[30.7]
[35.4]
[16.8]
[23.7]
[8.87]
2845
3016
3016
3016
3016
3016
2845
0.006
0.081
0.07
0.091
0.066
0.019
0.112
-0.770***
[-3.831]
-0.302**
[-2.309]
3.297***
[5.179]
-0.686***
[-4.379]
17.67***
[5.485]
2845
0.155
Estimation
Dataset Comparison
Sample
[4]
[3.82]
Total ICP Studies
Log(GDP)
Distance from Benchmark
Year, in Decades (1996)
Distance * Total ICPs
Constant
N
R-squared
[6]
OLS
PWT6.2 vs. PWT6.1
1970 to 1999
Abs(Difference in NIA GDP Growth) 1.144***
Abs(Difference in PWT-NIA
GDP Growth)
PWT Grade (D=1…A=4)
[5]
0.752***
[7.35]
-4.036***
[-18.0]
-2.319***
[-18.1]
***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust t-statistics in brackets.
Notes: The dependent variable is the absolute value of the difference in the log-level of GDP across PWT 6.2 and PWT 6.1. NIA
GDP is the GDP per capita, constant price, series from the World Development Indicators dataset (WDI), as reported by national
authorities. WDI 2002 data corresponds to PWT 6.1 and WDI 2005 corresponds to PWT 6.2. PWT-NIA GDP is a GDP percapita, constant price, series constructed using the Penn World Table's national income accounts datafiles. The variable PWT
grade can take on four values, 1-4, with 1 representing the worst PWT rating of D, and 4 representing the best PWT rating of A.
Total ICP studies refers to the number of ICP studies a country has participated in. GDP is total GDP and is measured at
purchasing power parity. The distance variable is calculated as the absolute difference between the year of the observation and
1996 (i.e., abs(t-1996)), and is divided by 10 to express the variable in decades. 1996 is the "Benchmark Year" used in
constructing both PWT 6.2 and PWT 6.1. The sample consists of the 104 countries in the “Long Run Sample” used in other
tables and figures.
41
Table 5. Explaining Revisions to Growth of GDP Per Capita Across Versions
Dependent variable
Abs(difference in growth rate of per capita GDP)
[1]
[2]
[3]
Estimation
Dataset Comparison
Sample
Abs(Difference in NIA GDP Growth)
Abs(Difference in PWT-NIA
GDP Growth)
PWT Grade (D=1…A=4)
Total ICP Studies
Log(GDP)
Distance from Benchmark
Year, in Decades (1996)
Constant
N
R-squared
[4]
[5]
OLS
PWT6.2 vs. PWT6.1
1970 to 1999
[6]
0.694***
[4.03]
[7]
[8]
0.537***
[3.09]
0.0865
[1.171]
0.826***
[26.22]
0.842***
[26.6]
-1.716***
[-17.8]
-0.678***
[-14.0]
-0.232***
[-4.82]
-0.784***
-0.576***
[-16.1]
[-11.2]
0.366*** 0.263**
[3.21]
[2.46]
2.403*** 0.432*** 6.767*** 4.409*** 22.12*** 2.355*** 16.96***
[26.0]
[6.32]
[22.6]
[23.6]
[17.5]
[15.4]
[12.9]
2845
3016
3016
3016
3016
3016
2845
0.017
0.747
0.093
0.058
0.096
0.003
0.11
-0.0538**
[-2.017]
-0.0967***
[-3.875]
-0.0130
[-0.209]
2.984***
[4.767]
2845
0.720
***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. Robust t-statistics in brackets.
Notes: The dependent variable is the absolute value of the difference in the growth rate of GDP across PWT 6.2 and PWT 6.1. NIA
GDP is the GDP per capita, constant price, series from the World Development Indicators dataset (WDI), as reported by national
authorities. WDI 2002 data corresponds to PWT 6.1 and WDI 2005 corresponds to PWT 6.2. PWT-NIA GDP is a GDP per-capita,
constant price, series constructed using the Penn World Table's national income accounts datafiles. The variable PWT grade can take
on four values, 1-4, with 1 representing the worst PWT rating of D, and 4 representing the best PWT rating of A. Total ICP studies
refers to the number of ICP studies a country has participated in. GDP is total GDP and is measured at purchasing power parity. The
distance variable is calculated as the absolute difference between the year of the observation and 1996 (i.e., abs(t-1996)), and is
divided by 10 to express the variable in decades. 1996 is the "Benchmark Year" used in constructing both PWT 6.2 and PWT 6.1.
The sample consists of the 104 countries in the “Long Run Sample” used in other tables and figures.
42
Table 6. Differences Between PWT Versions 6.2, 6.1, and 5.6
Annual Growth
Rates
Decadal Growth Rates
Long-Run
Growth Rate
1970 to 1990
1970s
1980s
1970-1990
GDP (RGDPCH)
PWT 6.2 - PWT 6.1
Avg GR (6.2)
Mean Difference
SD Difference
1.6%
0.0%
5.3%
2.2%
0.0%
1.7%
1.0%
0.0%
1.4%
1.6%
0.0%
1.0%
PWT 6.2 - PWT 5.6
Avg GR (6.2)
Mean Difference
SD Difference
1.6%
0.1%
5.9%
2.2%
-0.2%
1.9%
1.0%
0.5%
1.3%
1.6%
0.1%
1.1%
PRICES (P)
PWT 6.2 - PWT 6.1
Avg (6.2)
Mean Difference
SD Difference
62.5
2.7%
20.7%
70.4
3.3%
23.6%
67.4
2.4%
14.5%
65.0
2.9%
18.6%
PWT 6.2 - PWT 5.6
Avg (6.2)
Mean Difference
SD Difference
62.5
-10.2%
27.2%
70.4
-9.8%
31.5%
67.4
-5.1%
28.9%
65.0
-7.5%
29.5%
2000
100
100
100
Obs
Source: Authors’ calculations based on PWT 5.6, 6.1 and 6.2.
43
Table 7. Ramey and Ramey (1995) replication results
Dependent variable
Dataset substitution
Sample
PWT grades included
Standard deviation of
growth rates
N
growth of per capita GDP
[1]
[2]
[3]
[4]
[5]
None
Original
All
None
Balanced
All
PWT6.2
Balanced
All
None
Balanced
A,B,C
PWT6.2
Balanced
A,B,C
-0.177***
[2.426]
-0.151*
[1.821]
-0.074
[0.94]
-0.107
[1.254]
-0.132
[1.460]
2208
1776
1776
1608
1608
***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. t-statistics in brackets.
Notes: This table presents a replication exercise of Ramey and Ramey (1995) by updating their original PWT 5.6 GDP
data with PWT 6.2 data. The equation estimated is: y it   i  X it   it , where y is the log-level of per-capita GDP,
 i is the standard deviation of  it in i across t, and X is a vector of controls including initial investment share of GDP,
initial population growth rate, initial human capital, initial per-capita GDP, lagged GDP, and several time trend and
dummy variables. All specifications replicate Table 1, equation (1), in Ramey and Ramey (1995). For presentation
purposes we focus only on the key parameter - variation of growth rate - and omit other parameter estimates.
Specification [1] replicates the main result in Ramey and Ramey (1995) using their original data. Specification [2]
presents the same result using PWT 5.6 after dropping observations to balance the data with available observations in
PWT 6.2. Specification [3] presents our main result when we replace the balanced sample using PWT 5.6 with PWT
6.2. Specification [4] replicates [2] when countries with quality grading "D" are dropped. Specification [5] replicates [3]
when countries with quality grading "D" are dropped.
Source: Authors’ calculations based on Ramey and Ramey (1995).
44
Dependent variable
Table 8. Jones and Olken (2005) replication results
Annual growth rate
Hypothesis
Leader deaths affect growth (Wald p-value)
[2]
[3]
[4]
[1]
Dataset substitution
Sample
PWT grades included
None
Original
All
None
Balanced
All
Years after leader's death
t
1
2
Number of leader deaths
PWT6.2
Balanced
A,B,C
All leaders
0.093*
0.218
0.203
52
0.012**
0.028**
0.039**
45
0.110
0.130
0.141
45
0.039**
0.054*
0.031**
47
All leaders, tenure >=2 Years
0.039**
0.04**
0.011**
0.087*
0.140
0.021**
0.076*
0.102
0.023**
42
42
36
0.019**
0.016**
0.028**
29
0.032**
0.049**
0.100*
26
Autocrats
0.199
0.356
0.314
26
0.012**
0.011**
0.026**
21
0.304
0.227
0.183
21
0.326
0.415
0.370
20
Democrats
0.044**
0.092*
0.134
20
0.236
0.291
0.359
19
0.039**
0.069*
0.113
19
Years after leader's death
0
1
2
Number of leader deaths
None
Balanced
A,B,C
0.054*
0.122
0.133
52
Years after leader's death
0
1
2
Number of leader deaths
PWT6.2
Balanced
All
0.057*
0.085*
0.067*
57
Years after leader's death
0
1
2
Number of leader deaths
[5]
0.460
0.552
0.432
22
0.049**
0.068*
0.049**
36
***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. t-statistics in brackets.
Notes: This table presents a replication exercise of Jones and Olken (2005) by updating their original PWT 6.1 GDP data with
PWT 6.2 data The equation estimated is: g it   z PRE z   z POSTz   i   t   it , where g is growth of per-capita GDP and
vi and vj are fixed country and time effects. For each leader death at t, there are location-specific time dummy variables equal to 1
in one of t-5, t-4,…,t-1, t+1, t+2,…,t+5 and equal to 0 otherwise. These vectors of dummy variables are denoted PRE and POST.
Results are Wald p-values of the joint test that PREz
 POSTz . All specifications replicate results from Tables III and V in
Jones and Olken (2005). Specification [1] replicates the main result in Jones and Olken (2005) using their original data.
Specification [2] presents the same result using PWT 6.1 after dropping observations to balance the data with available
observations in PWT 6.2. Specification [3] presents our main result when we replace the balanced sample using PWT 6.1 with
PWT 6.2. Specification [4] replicates [2] when countries with quality grading "D" are dropped. Specification [5] replicates [3]
when countries with quality grading "D" are dropped. ***, **, and * indicate significance at the 1%, 5%, and 10% levels,
respectively.
Source: Authors’ calculations based on Jones and Olken (2005).
45
Table 9. Hausmann, Pritchett and Rodrik (2005) replication results
PWT 6.1
(Hausmann,
Pritchett, and
Rodrik)
PWT 6.2
PWT 6.1
(Hausmann,
Pritchett, and
Rodrik)
PWT 6.2
PWT 6.1
PWT 6.2
(Hausmann,
Pritchett, and
Rodrik)
ARG
1963
ARG
1963
IDN
1967
IDN 1967 PAK
1979
PAK 1977
ARG
1990
ARG
1990
IDN
1987
IDN 1985 PAN
1959
PAN 1959
AUS
1961
IND
1982
PAN
1975
PAN 1975
No break
No break
BEL
1959
BEL
1959
IRL
1958
IRL 1958 PER
1959
PER 1959
BRA
1967
BRA
1968
IRL
1985
IRL 1986
PHL 1970
No break
BWA 1984
IRN 1966 PNG
1987
No break
No break
No break
CAN
1962
CAN
1961
ISR
1957
ISR 1957 POL
1992
POL 1992
CHL
1986
CHL
1986
ISR 1989
PRT 1959
No break
No break
CHN
1978
CHN
1977
ISR
1967
ISR 1967 PRT
1985
PRT 1985
CHN
1990
CHN
1991
JOR
1973
PRY
1974
PRY 1974
No break
CMR
1972
JPN
1958
JPN 1958 ROM
1979
ROM 1971
No break
CMR 1978
JPN 1984 RWA
1975
RWA 1975
No break
No break
COG
1969
COG
1968
KOR
1962
KOR 1963 SGP
1969
SGP 1967
COG
1978
COG
1976
KOR
1984
KOR 1984
SGP 1987
No break
COL
1967
COL
1967
LKA
1979
SYR
1969
SYR 1969
No break
DNK
1957
DNK
1957
LKA 1958 SYR
1974
No break
No break
DOM
1969 DOM 1969
LSO
1971
LSO 1971 SYR
1989
SYR 1991
DOM
1992 DOM 1991
LSO 1992 TCD
1973
No break
No break
DZA
1975
MAR
1958
MAR 1958 THA
1957
THA 1957
No break
ECU
1970
ECU
1970
MLI
1972
MLI 1973 THA
1986
THA 1986
EGY
1976
EGY
1975 MUS
1971
MUS 1970 TTO
1975
No break
ESP
1959
ESP
1959 MUS
1983
MUS 1983 TUN
1968
TUN 1968
ESP
1984
ESP
1984 MWI
1970
MWI 1970 TWN
1961
TWN 1961
ETH
1988 MWI
1992
MWI 1990
TWN 1985
No break
No break
FIN
1958
FIN
1958 MYS
1970
MYS 1967
TZA 1992
No break
FIN
1967
MYS
1988
MYS 1988 UGA
1977
No break
No break
GAB
1969
NER 1974 UGA
1989
No break
No break
No break
FIN
1992
FIN
1992 NGA
1957
URY
1974
URY 1974
No break
GBR
1982
GBR
1982 NGA
1967
NGA 1968 URY
1989
URY 1988
GHA
1965
GHA
1965
NIC
1960
NIC 1960 USA
1961
USA 1961
GNB
1969
NOR
1991
NOR 1992
VEN 1971
No break
No break
GNB
1988
NZL
1957
NZL 1957
ZMB 1963
No break
No break
HND
1974
PAK
1962
PAK 1961 ZWE
1964
ZWE 1967
No break
No break
HTI
1990
Notes: This table compares growth breaks in per capita GDP (RGDPCH series) obtained from using the
original PWT 6.1 in Hausmann, Pritchett, and Rodrik (2005) with those obtained using PWT 6.2. Growth
breaks are reproduced using Hausmann, Pritchett, and Rodrik (2005) original Gauss code. No break indicates
an inconsistency between PWT 6.1 and PWT 6.2 in obtaining growth break dates that are more than three
years apart for each country considered. Hausmann, Pritchett, and Rodrik (2005) identified an additional break
for BWA in 1969 using PWT 6.1, but data for Botswana does not go farther back than 1970 in PWT 6.2.
There are 35 cases (out of 100; 35 percent) where such inconsistencies are detected.
Source: Authors’ calculations based on Hausmann, Pritchett, and Rodrik (2005).
46
Table 10. Miguel, Satyanath and Sergenti (2004) replication results
log (civil conflict)
Dependent variable
[1]
[2]
[3]
[4]
[5]
[6]
Estimation
Dataset substitution
Sample
PWT grades included
OLS
None
Original
All
OLS
None
Balanced
All
OLS
PWT
Balanced
All
2SLS
None
Original
All
2SLS
None
Balanced
All
2SLS
PWT
Balanced
All
Economic Growth Rate, t
-0.145
[0.767]
-0.144
[0.727]
-0.327*
[1.668]
-0.383
[0.276]
-0.284
[0.217]
-1.985
[1.351]
Economic Growth Rate, t-1
0.071
[-0.368]
0.079
[-0.395]
0.006
[-0.029]
-2.139**
[2.078]
-2.078**
[2.096]
-2.307**
[2.327]
743
0.52
724
0.50
724
0.49
743
0.39
724
0.38
724
0.29
N
R-squared
***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. t-statistics in brackets.
Notes: This table presents a replication exercise of Miguel, Satyanath and Sergenti (2004) by updating their original PWT 5.6
GDP data with PWT 6.2 data. The estimated equation is: conflict it   i  X it '    0 growthit   1 growthit 1   i yeart   it ,
with growthit instrumented by rainfallit and rainfallit-1 in the 2SLS specifications. X contains the log of per-capita GDP in
1979, a lagged democracy index (Polity IV), ethnolinguistic fractionalization, religious fractionalization, an oil-exporter
dummy variable, an index of topography (mountains), and the log of lagged population. All specifications replicate table 4,
equations 3 and 5, in Miguel, Satyanath and Sergenti (2004). For presentation purposes we focus only on the key parameter current and lagged growth rates - and omit other parameter estimates. Specification [1] replicates the main OLS result in
Miguel, Satyanath and Sergenti (2004) using their original data. Specification [2] presents the same result using PWT 5.6
after dropping observations to balance the data with available observations in PWT 6.2. Specification [3] presents our main
result when we replace the balanced sample using PWT 5.6 with PWT 6.2. Specifications [4], [5], [6] replicate specifications
[1], [2], [3], respectively, using 2SLS, using rainfall as an instrument for current and lagged economic growth.
Source: Authors’ calculations based on Miguel, Satyanath, and Sergenti (2004).
47
Table 11. Aghion, Howitt and Mayer-Foulkes (2005) replication results
log (growth gap)
Dependent variable
[1]
[2]
[3]
[4]
[5]
Dataset substitution
Sample
PWT grades included
None
Original
All
None
Balanced
All
PWT6.2
Balanced
All
None
Balanced
A,B,C
PWT6.2
Balanced
A,B,C
Financial development
-0.015
[0.93]
-0.015
[0.94]
0.011
[0.93]
-0.009
[0.60]
0.013
[1.21]
Financial development
* Initial GDP gap (1960)
-0.061***
[5.35]
-0.061***
[4.84]
-0.041***
[4.16]
-0.048***
[3.87]
-0.031**
[3.25]
Initial GDP gap (1960)
1.507***
[3.14]
1.505***
[2.83]
0.402
[1.02]
1.031*
[1.95]
0.090
[0.25]
71
0.51
60
0.34
60
0.37
57
0.31
57
0.37
24.70
37 out of 71
24.83
37 out of 71
9.76
65 out of 71
21.61
48 out of 71
2.92
71 out of 71
N
R-squared
Implied convergence threshold,   y /  fy
Number of countries above threshold
Instruments: legal origins, legal origins * initial GDP gap
Conditioning set: EMPTY
***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. t-statistics in brackets.
Notes: This table presents a replication exercise of Aghion, Howitt, and Mayer-Foulkes (2005) by updating their original
GDP data obtained from Levine, Loayza and Beck (2000) dataset with PWT 6.2 data. The equation estimated is:
gi  g1   0   f Fi   y ( yi  y1 )   fy Fi ( yi  y1 )   i , where yi-y1 is the gap in output between country i and the country at the
technology frontier, gi-g1 is the gap in per capita GDP growth, and F is the level of financial development. All
specifications replicate table 1, equation (1), in Aghion, Howitt, and Mayer-Foulkes (2005). Specification [1] replicates the
main result in Aghion, Howitt, and Mayer-Foulkes (2005) using their original data. Specification [2] presents the same
result using Levine, Loayza and Beck's data after dropping observations to balance the data with available observations in
PWT 6.2. Specification [3] presents our new result when we replace the balanced sample using Levine, Loayza and Beck's
data with PWT 6.2 data. Specification [4] replicates [2] when countries with quality grading "D" are dropped. Specification
[5] replicates [3] when countries with quality grading "D" are dropped.
Source: Authors’ calculations based on Aghion, Howitt, and Mayer-Foulkes (2005).
48
Table 12. Mankiw, Romer, and Weil (1992) replication results
Dependent variable
Log of GDP per capita in 1985
[1]
[2]
[3]
[4]
[5]
Dataset substitution
MRW sample
Sample
PWT grades included
None
Non-Oil
Original
All
None
Non-Oil
Balanced
All
PWT6.2
Non-Oil
Balanced
All
None
Non-Oil
Balanced
A,B,C
PWT6.2
Non-Oil
Balanced
A,B,C
log(I/GDP)-log(n+g+d)
0.738***
[5.96]
0.713***
[5.21]
0.431***
[3.22]
0.754***
[5.28]
0.430***
[2.99]
log(school) –log(n+g+d)
0.657***
[9.07]
0.622***
[7.92]
0.711***
[6.98]
0.639***
[7.59]
0.761***
[6.86]
N
R-squared
98
0.78
83
0.79
83
0.73
77
0.78
77
0.71
Solow restriction p-value
0.654
0.653
0.210
0.584
0.166
***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively. t-statistics are in
brackets.
Notes: This table presents a replication exercise of Mankiw, Romer, and Weil (1992) by
updating their original PWT 4.0 GDP and I/GDP data with PWT 6.2 data. The equation
estimated is:
ln(Y / L)   0   1 [ln( I / GDP )  ln(n  g   )]   2 [ln( school )  ln(n  g   )]   ,
where Y/L is GDP per worker, I/GDP is the investment share of GDP, n is the population growth
rate, g is the technology growth rate,  is the rate of capital depreciation, and school is average
number of working age adults in school from 1960-1985. All specifications replicate the Table 2
restricted regressions in Mankiw, Romer, and Weil (1992). The "Solow Restriction" is the
restriction that  1   2 . Specification [1] replicates the main result in Mankiw, Romer, and Weil
(1995) using their original data. Specification [2] presents the same result using PWT 4.0 for the
sample of non-oil countries after dropping observations to balance the data with available
observations in PWT 6.2. Specification [3] presents our main result when we replace the
balanced sample using PWT 4.0 with PWT 6.2. Specification [4] replicates [2] when countries
with quality grading "D" are dropped. Specification [5] replicates [3] when countries with quality
grading "D" are dropped.
Source: Authors’ calculations based on Mankiw, Romer, and Weil (1992).
49
ISO
Country
Appendix Table 1. Countries in the long-run sample
GDP Differences between PWT 6.2 and PWT 6.1 Benchmarks
Prices
ARG
AUS
AUT
BDI
BEL
BEN
BFA
BOL
BRA
BRB
BWA
CAN
CHE
CHL
CHN
CIV
CMR
COG
COL
COM
CPV
CRI
DNK
DOM
DZA
ECU
EGY
ESP
ETH
FIN
FJI
FRA
GAB
GBR
GER
GHA
GIN
GMB
GNB
GNQ
GRC
GTM
HKG
HND
HUN
IDN
IND
IRL
IRN
Argentina
Australia
Austria
Burundi
Belgium
Benin
Burkina Faso
Bolivia
Brazil
Barbados
Botswana
Canada
Switzerland
Chile
China
Cote d'Ivoire
Cameroon
Congo, Republic of
Colombia
Comoros
Cape Verde
Costa Rica
Denmark
Dominican Republic
Algeria
Ecuador
Egypt
Spain
Ethiopia
Finland
Fiji
France
Gabon
United Kingdom
Germany
Ghana
Guinea
Gambia, The
Guinea-Bissau
Equatorial Guinea
Greece
Guatemala
Hong Kong
Honduras
Hungary
Indonesia
India
Ireland
Iran
26.69
26.77
25.99
22.14
26.14
22.61
22.92
23.77
27.77
22.07
22.94
27.24
25.96
25.74
28.9
24.12
24.09
22.46
26.13
20.55
21.02
23.94
25.6
24.45
25.8
24.71
26.3
27.23
24.33
25.28
22.09
27.89
23.4
27.86
28.27
23.82
23.64
20.66
20.48
20.36
25.61
24.46
25.88
23.27
25.32
27.43
28.35
24.89
26.54
2%
8%
-6%
5%
2%
8%
-2%
4%
30%
10%
14%
6%
-1%
-5%
25%
12%
19%
-8%
3%
14%
-11%
0%
4%
-2%
4%
-14%
4%
1%
12%
4%
3%
0%
4%
0%
-4%
-3%
15%
48%
23%
78%
5%
0%
0%
0%
-1%
-5%
0%
7%
-17%
GDP
Growth
-0.08%
0.06%
-0.02%
0.84%
0.02%
0.10%
-0.32%
0.17%
-0.30%
-2.29%
-0.54%
-0.03%
0.20%
-0.34%
1.95%
1.11%
-0.03%
-1.48%
0.07%
0.79%
-0.73%
0.17%
0.03%
-0.31%
-0.30%
0.21%
0.41%
0.13%
1.55%
-0.32%
-0.22%
0.06%
-1.65%
0.12%
-0.02%
0.61%
-0.42%
-0.03%
-3.30%
7.18%
-0.14%
-0.23%
-0.14%
0.23%
0.08%
-0.48%
-0.01%
-0.33%
-0.24%
I/GDP
-9%
-1%
-7%
-30%
-5%
16%
3%
6%
-7%
-68%
2%
0%
4%
11%
49%
5%
-33%
16%
7%
56%
-11%
-42%
-6%
-18%
-9%
17%
17%
-7%
-16%
10%
-6%
-7%
-61%
-5%
4%
-4%
-28%
26%
-25%
-30%
-4%
-7%
-2%
1%
-3%
9%
-10%
13%
56%
Grade
C/GDP
-1%
-5%
-3%
2%
2%
-9%
13%
0%
3%
-15%
-7%
1%
2%
-5%
-16%
-4%
-7%
9%
1%
-9%
-7%
17%
-8%
-3%
-12%
6%
2%
-11%
5%
-21%
10%
-12%
-43%
-15%
-1%
17%
-7%
12%
-36%
-28%
-1%
-1%
1%
1%
-20%
-9%
-9%
-8%
-19%
2
3
5
0
6
2
0
2
3
2
3
4
2
2
0
3
3
2
3
0
0
1
5
1
0
2
2
5
2
4
1
6
1
6
1
0
1
0
0
0
4
1
3
1
5
2
4
5
4
B
A
A
C
A
C
C
C
C
C
C
A
A
B
C
C
C
C
C
D
D
C
A
C
D
C
C
B
C
A
C
A
C
A
B
C
C
C
D
D
B
C
A
C
C
C
C
A
C
50
ISO
Appendix Table 1 (contd.) . Countries in the long-run sample
Country
GDP Differences between PWT 6.2 and PWT 6.1 Benchmarks
Prices
ISL
ISR
ITA
JAM
JOR
JPN
KEN
KOR
LKA
LSO
LUX
MAR
MDG
MEX
MLI
MOZ
MRT
MUS
MWI
MYS
NAM
NER
NGA
NIC
NLD
NOR
NPL
NZL
PAK
PAN
PER
PHL
PNG
PRT
PRY
ROM
RWA
SEN
SLV
SWE
SYR
TCD
TGO
THA
TTO
TUN
TUR
TZA
UGA
URY
USA
VEN
ZAF
ZMB
ZWE
Iceland
Israel
Italy
Jamaica
Jordan
Japan
Kenya
Korea, Republic of
Sri Lanka
Lesotho
Luxembourg
Morocco
Madagascar
Mexico
Mali
Mozambique
Mauritania
Mauritius
Malawi
Malaysia
Namibia
Niger
Nigeria
Nicaragua
Netherlands
Norway
Nepal
New Zealand
Pakistan
Panama
Peru
Philippines
Papua New Guinea
Portugal
Paraguay
Romania
Rwanda
Senegal
El Salvador
Sweden
Syria
Chad
Togo
Thailand
Trinidad &Tobago
Tunisia
Turkey
Tanzania
Uganda
Uruguay
USA
Venezuela
South Africa
Zambia
Zimbabwe
22.5
25.43
27.8
23.2
23.52
28.72
24.32
27.19
24.89
21.78
23.44
25.4
23.18
27.21
22.96
23.44
21.93
23.41
22.8
26.03
22.81
22.71
25.47
23.39
26.6
25.59
24.1
24.97
26.46
23.64
25.35
26.23
23.68
25.75
23.99
25.55
22.44
23.19
23.97
25.99
24.09
22.56
22.08
26.75
23.24
24.73
26.56
23.44
23.67
24.24
29.74
25.83
26.45
22.74
24.28
2%
-2%
1%
-18%
-24%
5%
0%
9%
-2%
6%
-5%
0%
-3%
11%
-4%
89%
16%
-4%
2%
2%
-26%
-3%
6%
18%
-1%
-1%
1%
-1%
3%
2%
-9%
-2%
28%
-5%
-10%
22%
3%
-3%
0%
1%
58%
22%
10%
1%
-13%
-1%
27%
-10%
-46%
-3%
0%
1%
7%
3%
-25%
GDP
Growth
0.03%
-0.06%
-0.09%
0.17%
-2.21%
-0.07%
-1.15%
0.16%
1.49%
1.39%
-0.09%
-0.08%
-0.38%
0.13%
1.05%
1.76%
1.73%
0.09%
-0.06%
0.94%
0.00%
1.25%
0.87%
0.92%
-0.20%
0.06%
-0.30%
0.16%
-0.13%
0.92%
-0.22%
0.08%
2.69%
-0.07%
-0.02%
0.05%
-0.22%
0.10%
0.60%
-0.05%
-0.91%
0.38%
0.04%
0.10%
-0.65%
-0.09%
0.02%
1.83%
-1.56%
0.28%
-0.06%
1.35%
0.42%
-0.07%
-0.34%
I/GDP
-5%
-11%
-1%
-8%
8%
-1%
10%
2%
27%
13%
8%
-9%
33%
-2%
14%
17%
110%
-8%
-23%
-2%
-36%
-10%
-16%
-21%
-1%
-11%
4%
-3%
6%
-9%
-7%
-7%
-38%
-4%
1%
-28%
-31%
-20%
9%
0%
-34%
-8%
44%
1%
91%
21%
-14%
-77%
15%
10%
-2%
1%
-21%
62%
-36%
Grade
C/GDP
-7%
-9%
-5%
23%
11%
-8%
-1%
-3%
-10%
15%
-7%
1%
-9%
-6%
-7%
-8%
-2%
-4%
7%
-2%
1%
-14%
-3%
-11%
-6%
-27%
-1%
-15%
-7%
14%
-1%
-2%
-5%
-5%
18%
-15%
13%
13%
17%
-7%
-8%
6%
-7%
6%
-1%
2%
1%
24%
-6%
-1%
0%
1%
5%
25%
6%
2
2
6
3
1
6
5
5
4
0
5
3
3
3
3
0
0
2
4
2
0
0
3
0
6
4
2
1
4
2
2
5
0
4
1
2
1
3
1
3
2
0
0
3
2
3
3
3
0
3
6
2
0
4
3
B
B
A
C
C
A
C
B
C
D
A
C
C
C
C
D
C
C
C
C
D
D
C
C
A
A
C
B
C
C
C
C
D
B
C
C
C
C
C
A
C
D
D
C
C
C
C
C
D
B
A
C
C
C
C
51
Appendix Table 2. Replication exercise results summary
Author
Journal Year
Tables
Replicated
Replication Exercise Results Summary
Results not robust to dataset
Aghion, Howitt and
Mayer-Foulkes
QJE
2005
Hausmann, Rodrik,
and Pritchett
JEG
2005
Jones and Olken
QJE
2005
Ramey and Ramey
AER
1995
Table 1,
Cols 1-4
Initial GDP gap becomes insignificant in many
exercises. Implied convergence thresholds
become implausibly low.
Table 1
Discrepancies identified in 35 out of 100 growth
acceleration cases between PWT 6.2 and PWT
6.1.
Many differences across exercises. Most
Tables 3, 5 pronounced are changes in significance of
Democrats' and Autocrats' deaths.
Table 1
Standard deviation of growth rates coefficient
becomes weakly significant in one exercise and
insignificant in three.
Results robust to dataset changes:
Acemoglu, Johnson,
Robinson and
Thaicharoen
JME
Table 2,
2003
Rows 1-4
Barro
JPE
1999
Minor changes to parameter magnitudes and
significance levels.
Minor changes to parameter magnitudes and
Table 1, Col significance levels.
2
Burnside and Dollar
AER
2000 Tables 1-3
Some changes in significance levels, but
parameter magnitudes remain largely unchanged.
DeLong and Summers
QJE
1991 Table 1
Some changes in parameter magnitudes, but
significance levels remain largely unchanged.
2004 Tables 5, 7
Minor changes to parameter magnitudes and
significance levels.
Demirguc-Kunt,
Laeven and Levine
Easterly, Kremer
Pritchett, and
Summers
Mankiw, Romer and
Weil
Miguel, Satyanath and
Sergenti
Sachs and Warner
JMCB
Table 5,
Cols 2, 4
Minor changes to parameter magnitudes and
significance levels.
JME
1993
QJE
Table 2,
1992
Cols 1, 2
I/GDP coefficient halves in magnitude, nearly
violating CRS Solow restriction in some
exercises. Otherwise estimates are robust.
JPE
Table 4,
2004
Cols 1-7
Instruments seem to reduce or eliminate
parameter estimate differences when using PWT
6.2 data vs. their data.
1995
Table 11,
Col 7
Some changes in significance levels, but
parameter magnitudes remain largely unchanged.
52
Figure 1. Differences in 29-year average per capita GDP growth rates between PWT 6.2
and PWT 6.1
Differences in 29 Year Average Growth Rates
-.025
0
.025
PNG
CHN
TZA
MOZ
MRT
ETH LKA
VEN
NER
CIV
MLI
NIC
PAN
NGAMYS
BDI
GHA
SLV
LSO
COM
IND
ARG
LUX
USA
BRA
BWA
TTO
CPV
SYR
KEN
COG
BRB
UGA
GAB
JOR
average difference: .0013
sd of difference: .0109
average growth rate: .0156
GNB
20
22
24
26
Log of GDP in 1996
28
30
Notes: Twenty-nine-year average annual per capita GDP growth rates are computed using the RGDPCH series
for the period 1970–1999. Differences in growth rates between the two versions of PWT are calculated as GDP
growth from PWT 6.2 minus GDP growth from PWT 6.1. Average differences for 29-year average annual per
capita GDP growth rates are very close to zero (.0013) whereas the standard deviation is .0109. Average growth
rate is about 1.56 percent. The sample consists of the 104 countries in the “long-run sample” used in other
tables and figures.
Source: Authors’ calculations.
53
Differences in 29 Year Averages
0
.5
1
Figure 2. Differences in 29-year average prices between PWT 6.2 and PWT 6.1
MOZ
GNQ
SYR
GMB
GNB
COM
PNG
TCD NIC CMR
MRT
BWA GIN CIV
ETH
ROM
BRA
TUR
MEX
ARG
LUX
TZA PRY
TTO
ECU
JAM
NAM JOR ZWE
CPV
CHN
IND
USA
IRN
-.5
average % difference: 2.89
sd of % difference: 15.39
average level: 62.529
UGA
20
22
24
26
Log of GDP in 1996
28
30
Notes: Twenty-nine-year average prices are computed using the average of the P series for the period 1970–
1999. Differences in averages between the two versions of PWT are calculated as the log of average prices from
PWT 6.2 minus the log of average prices from PWT 6.1. Average differences for 29-year average prices are
very close to zero (2.89) whereas the standard deviation is 15.39. The average price level is 62.93, relative to a
price level of 100 for the United States. The sample consists of the 104 countries in the “long-run sample” used
in other tables and figures.
Source: Authors’ calculations.
54
Figure 3. Differences in annual per capita GDP growth between PWT 6.2 and PWT 6.1
20
22 24 26 28 30
Log of GDP in 1996
-.5
GDP Growth Rate Differences
-.25
0
.25
.5
.75
29 Year Panel
-.75
-.5
GDP Growth Rate Differences
-.25
0
.25
.5
.75
10 Year Panel
-.75
-.75
-.5
GDP Growth Rate Differences
-.25
0
.25
.5
.75
1 Year Panel
20
22 24 26 28 30
Log of GDP in 1996
20
22 24 26 28 30
Log of GDP in 1996
Notes: Differences in growth rates between the two versions of PWT are calculated as GDP growth from PWT
6.2 minus GDP growth from PWT 6.1. 1 Yr Panel: # of obs: 2880; avg. difference = .0008; sd of difference =
.0572; avg. growth rate = .0157; 10 Yr Panel: # of obs: 288; difference = .0010; sd of difference = .0174; avg.
growth rate = .0155; 29 Yr Panel: # of obs: 96; difference = .0010; sd of difference = .0108; avg. growth rate
=.0151. The sample of countries is constant across the three figures above, and consists of 96 of the 104
countries in the “long-run sample” used in other tables and figures. Eight countries were dropped because they
did not have data for all three decades. This sample is called the “decades sample.”
Source: Authors’ calculations.
55
Figure 4. GDP per capita: PWT 6.2, PWT 6.1, and PWT 5.6
6
8.5
8
7
9
8.5
8
9.5
9
China
9
Brazil
10
Barbados
1970
1980
1990
2000
1960
1970
1980
1990
2000
1960
1970
1980
1990
2000
1990
2000
India
7
7
6
7.5
6.5
7.5
8
7
8
7.5
Ghana
8.5
Egypt
1970
1980
1990
2000
1960
1970
1990
2000
Sri Lanka
1960
1970
1980
United States
7
9.5
7.5
10
10
8
10.5
11
8.5
Luxembourg
1980
10.5
1960
9.5
GDP Per Capita
1960
1960
1970
1980
1990
2000
PWT62
1960
1970
1980
1990
PWT61
2000
1960
1970
1980
1990
PWT56
Notes: Per-capita GDP is measured in 2000 international dollars. Because PPPs in each version of the Penn
World Table are relative to USA, and base years in each version of the PWT are different, re-basing requires
USA GDP deflators for the base years in each of the PWT versions.
Source: Authors’ calculations.
2000
56
Figure 5. Countries with respective PWT data quality grades
20
Number of Countries
40
60
80
100
100
47
19
0
13
A
B
Source: ICP benchmark studies and quality grades used in PWT 6.1.
C
D
57
Figure 6. Differences in 29-year average GDP growth by PWT quality grade
between PWT 6.2 and PWT 6.1
MOZ
.5
.5
GNQ
SYR
GMB
KOR
IRL GRC AUS
CAN
JPN
FIN
DNK
BELARG ITA
SWE
USA
ISRCHE
GER
LUX URY
PRT
CHL
AUT
PNG
TUR BRA CHN
GNB
TCDNICCMR ROM
GIN
COM MRT BWA
CIV
ETH
MEX
BRB
TGO
BEN
ZAF
LSO
NGA
BDI
BOL
DZA
EGY
GAB
COL
PAK
ZMB
FJI
RWA
MWIPAN
MYS
NPL
THA
IND
TUN
DOM
LKA PHL
BFA
NER
SEN
MDG
GHA
MLI
MUS
IDN
COGTZA
PRY PER
CPV
TTO ECU
IRN
JAM
0
0
ISL
-.5
JORZWE
NAM
-.5
Differences in 29 Year Averages
1
C or D
1
A or B
UGA
22
24
26
28
30
20
25
30
Log of GDP in 1996
Notes: Twenty-nine-year average annual per capita GDP growth rates are computed using the RGDPCH
series for the period 1970–1999. Differences in growth rates between the two versions of PWT are
calculated as GDP growth from PWT 6.2 minus GDP growth from PWT 6.1. The sample consists of the
104 countries in the “long-run sample” used in other tables and figures. A or B: average difference: -.0002;
sd of difference: .0015; average growth rate: .0233; countries: 29. C or D: average difference: .0019; sd of
difference: .0127; average growth rate: .0126; countries: 75.
Source: Authors’ calculations.
58
Figure 7. Evolution of differences in levels of prices across time between PWT 6.2
and PWT 6.1
Notes: Prices is the Price (P) series. Differences in levels between the two versions of PWT are calculated
as the log level of P from PWT 6.2 minus the log level of P from PWT 6.1. Each mean and standard
deviation is computed across countries for a given year. Sample of countries for each year includes
countries for which there is data for every year between 1970 and 1999. Vertical line denotes benchmark
year, 1996. The sample consists of the 104 countries in the “long-run sample” used in other tables and
figures.
Source: Authors’ calculations.
59
THESE APPENDIX FIGURES ARE NOT FOR PUBLICATION
Appendix Figure 1. Countries with respective number of ICP benchmark studies
40
43
Number of Countries
20
30
32
27
21
10
12
9
0
7
0
1
2
Source: ICP benchmark studies used in PWT 6.1.
3
4
5
6
60
Appendix Figure 2. 10-year average per capita GDP growth by decade
between PWT 6.2 and PWT 6.1
.1
.1
1970s
GNQ
.05
IND USA
ARGBRA
CHN
LUX
MOZ
NER
0
GHA
LUX
-.05
.05
GNQ
-.05
0
ROM
GNB
COG
CPV
GNB
CHN
IND
ARGBRA
USA
JOR
-.1
-.1
TCD
20
24
28
32
20
24
.1
GNQ
GNB
ARGBRA
IND USA
GNB
TZA
UGA
NGA
CHN
IND USA
ARGBRA
COG
-.1
ROM
24
LUX
-.05
LUX
0
32
CHN
GNQ
-.05
.05
.05
TCD
20
28
1990s
0
.1
1980s
-.1
Differences in 10 Year Average GDP Growth Rates
1960s
28
32
20
24
28
32
Log GDP in 2000
Notes: Ten-year average annual per capita GDP growth rates are computed using the RGDPCH series for
the period 1960–2000. Average annual growth rates are calculated as [log(RGDPCHt+10)- log(RGDPCHt)]/10. Differences in growth rates between the two versions of PWT are calculated as GDP growth from
PWT 6.2 minus GDP growth from PWT 6.1. 1960s: avg. difference = .0001; sd of difference = .0123; avg.
growth rate = .0279. 1970s: avg. difference = .0012; sd of difference = .0217; avg. growth rate = .0212.
1980s: avg. difference = -.0008; sd of difference = .0163; avg. growth rate = .0084. 1990s: avg. difference
= .0028; sd of difference = .0130; avg. growth rate = .0168. The sample of countries is constant across
decades, and consists of 96 of the 104 countries in the “long-run sample” used in other tables and figures.
Eight countries were dropped because they did not have data for all four decades. This sample is called the
“decades sample.”
Source: Authors’ calculations.
61
Appendix Figure 3. Evolution of differences in per capita GDP growth across time
between PWT 6.2 and PWT 6.1
Notes: GDP is the real per-capita GDP (RGDPCH) series. Differences in growth rates between the two
versions of PWT are calculated as the log difference of RGDPCH from PWT 6.2 minus the log difference
of RGDPCH from PWT 6.1. Each mean and standard deviation is computed across countries for a given
year. Sample of countries for each year includes countries for which there is data for every year between
1970 and 1999. Vertical line denotes benchmark year, 1996. The sample consists of the 104 countries in the
“long-run sample” used in other tables and figures.
Source: Authors’ calculations.